FACULDADE DE E NGENHARIA DA U NIVERSIDADE DO P ORTO

Application of Graph Neural Networks

in Road Traffic Forecasting for
Intelligent Transportation Systems

Ana Clara Moreira Gadelho

Mestrado em Engenharia Informática e Computação

Supervisor: Dr. Daniel Augusto Gama de Castro Silva

Second Supervisor: Dr. Rosaldo José Fernandes Rossetti

July 21, 2023

 Application of Graph Neural Networks in Road Traffic
Forecasting for Intelligent Transportation Systems

Ana Clara Moreira Gadelho

Mestrado em Engenharia Informática e Computação

Approved in oral examination by the committee:

Chair: Prof. Rui Camacho

External Examiner: Prof. Marco Veloso
Supervisor: Prof. Daniel Castro Silva

July 21, 2023

 i

This work is a result of project DynamiCITY: Fostering Dynamic Adaptation of Smart Cities
to Cope with Crises and Disruptions, with reference NORTE-01-0145-FEDER-000073, supported
by Norte Portugal Regional Operational Programme (NORTE 2020), under the PORTUGAL 2020
Partnership Agreement, through the European Regional Development Fund (ERDF).
Resumo

A previsão de tráfego é um aspecto crucial dos Sistemas de Transporte Inteligentes, uma vez
que tem o potencial de melhorar a mobilidade e eficiência do transporte nas cidades, reduzindo os
custos e minimizando o impacto ambiental. A tarefa de previsão de tráfego é um desafio complexo,
uma vez que envolve lidar com a rápida evolução e o dinamismo do tráfego, que é afectado por
vários factores, tais como acidentes, cortes de estradas, eventos sociais e meteorológicos. Para
além disso, o fluxo de tráfego é caracterizado por dependências espaciais e temporais, em que o
estado do tráfego num determinado local é influenciado por o tráfego noutros locais e o estado do
tráfego num determinado momento depende do tráfego passado.
Nos últimos anos, as Graph Neural Networks (GNNs) têm ganho cada vez mais atenção no
terreno de Deep Learning, demonstrando desempenho ao nível de estado da arte em várias apli-
cações. As GNNs são particularmente adequadas a problemas de previsão de tráfego, uma vez que
têm a capacidade de capturar tanto dependências espaciais como temporais nos dados.
Esta dissertação explora a utilização de GNNs na previsão de tráfego rodoviário. O estudo
fornece uma revisão da literatura existente sobre técnicas de previsão de tráfego e investiga o
potencial das GNNs no tratamento das relações complexas entre as condições de tráfego, tanto no
tempo como no espaço. A investigação tem como objetivo avaliar o desempenho das GNNs neste
problema e complementar esse estudo com a investigação do impacto da utilização de técnicas de
imputação de dados em falta e dos factores externos nos resultados da previsão. O estudo também
tem como objetivo avaliar a generalização dos modelos GNN em diferentes datasets, incluindo
datasets de benchmarking e o dataset de caso de uso do projeto DynamiCITY, que esta dissertação
integra.
A avaliação empírica dos modelos demonstra a sua eficácia no tratamento das relações com-
plexas entre as condições de tráfego no tempo e no espaço. Os resultados da investigação mostram
que as GNNs superam vários modelos baseados em séries temporais comummente utilizados na
literatura, em linha com outros trabalhos sobre este tema. As técnicas de imputação de dados em
falta também apresentaram melhores resultados, especialmente nos conjuntos de dados com mais
dados em falta. No entanto, a utilização de dados meteorológicos não melhorou os resultados
obtidos.
Em termos gerais, esta dissertação contribui para a crescente investigação sobre a utilização de
GNNs na previsão de tráfego e demonstra o seu potencial para melhorar os sistemas de transporte.

ii
Abstract

Traffic forecasting is a crucial aspect of Intelligent Transportation Systems, as it has the potential to
improve the mobility and efficiency of transportation in cities while reducing costs and minimizing
environmental impact. The task of traffic forecasting is a challenging one, as it involves predicting
the rapidly changing and dynamic nature of traffic, which is affected by various factors such as
accidents, road closures, social events, and weather. Additionally, traffic flow is characterized by
both spatial and temporal dependencies, where the state of traffic in one location is influenced by
traffic in other locations and the state of traffic at a particular time is dependent on past traffic
patterns.
In recent years, Graph Neural Networks (GNNs) have gained increasing attention in the field
of deep learning, demonstrating state-of-the-art performance in various applications. GNNs are
particularly well suited for traffic forecasting problems, as they have the ability to capture both
spatial and temporal dependencies in the data.
This dissertation explores the application of GNNs in road traffic forecasting. The study pro-
vides a review of existing literature on traffic forecasting techniques and investigates the potential
of GNNs in handling the complex relationships between traffic conditions in time and space. The
research aims to evaluate the performance of GNNs in this problem and complement that study by
investigating the impact of using missing data imputation techniques and external factors on the
prediction results. The study also aims to evaluate the generalizability of the GNN models across
different datasets, including benchmarking datasets and DynamiCITY’s use-case dataset.
The empirical evaluation of the models demonstrates their effectiveness in handling the com-
plex relationships between traffic conditions in time and space. The results of the research show
that GNNs outperform several commonly used time-series-based models in the literature, in line
with other works on the topic. The missing data imputation techniques also showed improved
results, especially on the datasets with more missing data. However, the use of weather data did
improve the results obtained.
Overall, this dissertation contributes to the growing body of research on the use of GNNs in
traffic forecasting and demonstrates their potential to improve transportation systems.

Keywords: traffic forecasting, intelligent transportation systems, graph neural networks, spatio-
temporal dependency, deep learning, missing data, external factors

iii
Acknowledgments

First, I would like to thank Professor Daniel Silva and Professor Rosaldo Rossetti, supervisors of
this dissertation, for their guidance and encouragement throughout the duration of this dissertation
process. Without their support in the most stressful phases this work could not be the way it is. I
would like to acknowledge the support and resources provided by LIACC and the DynamiCITY
research project. The access to computing facilities and data provided were crucial in conducting
the experiments and analysis for this research.
I also want to express my gratitude for my family, in particular my parents and brother, for
always believing in me and giving me encouragement for all the challenges I face.
Finally, I want to thank all my friends, that accompanied me throughout this five years and in
particular this last few months . Without them this journey would not have been the same.

Clara Gadelho

iv
Contents

1 Introduction 1
1.1 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Aim and Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Background 4
2.1 Intelligent Transportation Systems . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 The Traffic Forecasting Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2.1 Traffic Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.2 Categorization of Traffic Prediction Problems . . . . . . . . . . . . . . . 5
2.3 Graph Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3.1 General GNN Arquitecture . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3.2 Attention Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3 Related Work 12
3.1 Traffic Forecasting Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.1 Statistical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 Regression Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.1 Ensemble Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.2 Hybrid Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.3 Machine Learning and Deep Learning Techniques . . . . . . . . . . . . 14
3.2.4 Graph Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 Benchmarking Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4 Gap Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4 Methodological Approach 22
4.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2 Problem Formalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3 Implementation Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.4 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.4.1 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.4.2 Baselines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.5 Risk Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

5 Implementation 27
5.1 Technologies Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.2 Data Retrieval and Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . 28

v
CONTENTS vi

5.2.1 Traffic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

5.2.2 Weather Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.2.3 Event Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.3 Graph Dataset Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.3.1 Adjacency Matrix Generation . . . . . . . . . . . . . . . . . . . . . . . 37
5.3.2 Traffic Speed Normalization . . . . . . . . . . . . . . . . . . . . . . . . 38
5.3.3 Additional Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.3.4 Building Inputs and Targets . . . . . . . . . . . . . . . . . . . . . . . . 39
5.3.5 Train, Validation and Test Splitting . . . . . . . . . . . . . . . . . . . . 39
5.3.6 Dataloader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.4 Base Graph Neural Network Architecture . . . . . . . . . . . . . . . . . . . . . 41
5.4.1 Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.4.2 Network Hyperparameters . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.5 Handling Missing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.5.1 Missing data Occurrence and Characteristics . . . . . . . . . . . . . . . 45
5.5.2 Missing Data Handling Techniques . . . . . . . . . . . . . . . . . . . . 46
5.5.3 Comparing the techniques . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.5.4 Inclusion in the Data Processing Pipeline . . . . . . . . . . . . . . . . . 51
5.6 External Data Incorporation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.6.1 Weather Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.6.2 Event Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

6 Empirical Evaluation 55
6.1 Base Graph Neural Network Architecture . . . . . . . . . . . . . . . . . . . . . 56
6.1.1 Model Comparisons in Each Dataset . . . . . . . . . . . . . . . . . . . . 56
6.1.2 Effect of Hyperparameters . . . . . . . . . . . . . . . . . . . . . . . . . 59
6.1.3 Effect of the Length of the Training Set . . . . . . . . . . . . . . . . . . 67
6.2 Handling Missing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.3 Weather Data Incorporation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.4 Global Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.5 Results Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

7 Conclusions and Future Work 75

References 77
List of Figures

2.1 Grid and Graph-Based Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2 Traffic Prediction Categorization . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Comparison between a 2-D Convolution and a Graph Convolution . . . . . . . . 8
2.4 GNN traffic prediction formulation . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5 Spatio-temporal GNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.6 Attention Mechanism Representation . . . . . . . . . . . . . . . . . . . . . . . . 11

3.1 Hybrid CNN-RNN Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3.2 Benchmarking Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

4.1 Model Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

5.1 Benchmarking Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

5.2 METR-LA and PeMS-BAY Dataset Structure . . . . . . . . . . . . . . . . . . . 30
5.3 VCI Sensor Location . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.4 Event category destribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.5 Sensor Location Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.6 Description of the graph dataset generation . . . . . . . . . . . . . . . . . . . . 37
5.7 Structure of the GNN layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.8 Misisng data type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.9 Mean imputation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.10 LOCF imputation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.11 Interpolation imputation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.12 MICE imputation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.13 Data processing pipeline with missing data imputation . . . . . . . . . . . . . . 51
5.14 Effect of weather in traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.15 VCI Sensors and direction of traffic flow . . . . . . . . . . . . . . . . . . . . . . 53
5.16 Effect of events in traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

vii
List of Tables

2.1 Notation used to describe the GNN architecture . . . . . . . . . . . . . . . . . . 9

3.1 Existing GNN Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3.2 List of traffic datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

5.1 Different Traffic Datasets Description . . . . . . . . . . . . . . . . . . . . . . . 29

5.2 VCI Dataset Field Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.3 Weather Dataset Field Description . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.4 Events Dataset Field Description . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.5 Temporal and Spatial Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.6 Presence of missing values in the dataset . . . . . . . . . . . . . . . . . . . . . . 46
5.7 Empirical evaluation results of the different imputation techniques . . . . . . . . 50

6.1 Identification of the names used for the different GNN arquitectures . . . . . . . 55
6.2 Comparison of the different models on the PeMS-BAY dataset . . . . . . . . . . 56
6.3 Comparison of the different models on the METR-LA dataset . . . . . . . . . . . 57
6.4 Comparison of the different models on the VCI dataset . . . . . . . . . . . . . . 58
6.5 Comparison of the different optimizers used with the Cheb-GRUAtt on the PeMS-
BAY dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
6.6 Comparison of the different optimizers used with the Cheb-GRUAtt on the METR-
LA dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.7 Comparison of the different optimizers used with the Cheb-GRUAtt on the VCI
dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.8 Comparison of the different optimizers used with the GAT-CNNAtt on the PeMS-
BAY dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.9 Comparison of the different optimizers used with the GAT-CNNAtt on the METR-
LA dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.10 Comparison of the different optimizers used with the GAT-CNNAtt on the VCI
dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.11 Results of different learning rates with 100 epochs for Cheb-GRUAtt on PeMS-
BAY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.12 Results of different learning rates with 100 epochs for Cheb-GRUAtt on METR-
LA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.13 Results of different learning rates with 100 epochs for Cheb-GRUAtt on VCI . . 62
6.14 Results of different learning rates and with 100 epochs for GAT-CNNAtt on PeMS-
BAY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.15 Results of different learning rates and with 100 epochs for GAT-CNNAtt on METR-
LA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.16 Results of different learning rates with 100 epochs for GAT-CNNAtt on VCI . . 63

viii
LIST OF TABLES ix

6.17 Abbreviations for time features to use in the experiments’ tables . . . . . . . . . 64

6.18 Results of Cheb-GRUAtt with the use of different combinations of time features
in PeMS-BAY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.19 Results of Cheb-GRUAtt with the use of different combinations of time features
in METR-LA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.20 Results of Cheb-GRUAtt with the use of different combinations of time features
in VCI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.21 Results of GAT-CNNAtt with the use of different combinations of time features in
PeMS-BAY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.22 Results of GAT-CNNAtt with the use of different combinations of time features in
METR-LA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.23 Results of GAT-CNNAtt with the use of different combinations of time features in
VCI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.24 Results of Cheb-GRUAtt with different numbers of historical features and output
predictions on METR-LA dataset . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.25 Results of GAT-CNNAtt with different numbers of historical features and output
predictions on METR-LA dataset . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.26 Results of different dataset lengths . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.27 Results of the missing data imputation techniques in PeMS-BAY . . . . . . . . . 69
6.28 Results of the missing data imputation techniques in METR-LA . . . . . . . . . 69
6.29 Results of the missing data imputation techniques in VCI . . . . . . . . . . . . . 70
6.30 Abbreviations for weather features to use in the experiments’ tables . . . . . . . 71
6.31 Results with the use of different combinations of weather features in PeMS-BAY 71
6.32 Results with the use of different combinations of weather features in METR-LA . 71
6.33 Results with the use of different combinations of weather features in VCI . . . . 71
6.34 Comparison of the best-performing models with baselines in PeMS-BAY . . . . 72
6.35 Comparison of the best-performing models with baselines in METR-LA . . . . . 72
6.36 Comparison of the best-performing models with baselines in VCI . . . . . . . . 73
Listings

5.1 Weather tool usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

5.2 Events tool usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

x
xi
Acronyms and Abbreviations xii

Acronyms and Abbreviations

AI Artificial Intelligence
Agg Aggregation
Att Attention
ARIMA Autoregressive Integrated Moving Average
ASTGCN Attention Based Spatial-Temporal Graph Convolutional Network
ChebConv Chebyshev Spectral Graph Convolution
CNN Convolutional Neural Network
DL Deep Learning
DNN Deep Neural Network
DSTGCN Dynamic Spatial-Temporal Graph Convolutional Network
FNN Feedforward Neural Network
GAT Graph Attention Network
GCN Graph Convolution Network
GMAN Graph Multi-Attention Network
GNN Graph Neural Network
GRU Gated Recurrent Unit
GTCN Temporal Graph Convolutional Network
ITS Intelligent Transportation Systems
LOCF Last Observation Carried Forward
LSTM Long Short-term Memory
MAE Mean Absolute Error
MAPE Mean Absolute Percentage Error
MICE Multivariate Imputation by Chained Equations
ML Machine Learning
MLP Multilayer Perceptron
NN Neural Network
NOCB Next Observation Carried Backwards
N/A Non Applicable
PeMS Caltrans Performance Measure Systems
RMSE Root Mean Squared Error
RNN Recurrent Neural Network
SLC Structure Learning Convolution
STGCN Spatio-Temporal Graph Convolutional Network
ST-UNet Spatio-Temporal U-Network for Graph-structured Time Series Modeling
SVM Support Vector Machine
T-GCN Temporal Graph Convolutional Network
VAR Vector Auto Regressive Model
VCI Via de Cintura Interna (Porto’s Ring Road)
WHO World Health Organization
Chapter 1

Introduction

1.1 Context
With the acceleration of urbanization and the rapid growth of urban population, great pressure is
being placed on urban traffic management (United Nations, 2017). Expanding cities face numer-
ous challenges related to transportation, including increased air pollution and worsening traffic
congestion (W. Jiang & Luo, 2022). According to the World Health Organization (WHO, 2022),
the transportation sector leads to a significant portion of air pollution, with more than seven million
people dying each year due to this crisis. With this in mind, great effort is nowadays being placed
into developing Intelligent Transportation Systems (ITS), which could significantly improve the
lives of residents in future cities (Shahid, Shah, Khan, Maple, & Jeon, 2021). In the context of
ITS, Traffic forecasting is seen as an essential step in improving the efficiency of the transporta-
tion system and alleviating transportation-related problems (J. Lu, Li, Li, & Al-Barakani, 2021).
This is a complex task since traffic is very dynamic and involves dependencies in terms of space
and time (W. Jiang & Luo, 2022). So, over the years, researchers have been iterating over differ-
ent ways of improving traffic predictions. This dissertation is being prepared within the scope of
an academic research project: DynamiCITY. DynamiCITY aims to create a virtual environment
for the design and implementation of solutions for mobility within smart cities. It consists of an
open-data platform and a multi-agent-based decision-support system that allows researchers and
practitioners to explore the relationship between the city and its transport system, focusing on cli-
mate neutrality, safety, and accessibility. The DynamiCITY is being developed as a collaboration
between three research laboratories of FEUP: LIACC, CITTA, and SYSTEC. It will represent an
important R&D infrastructure available to the whole research community.

1.2 Aim and Goals

The main aim of this dissertation is to investigate the use of machine learning models for road
traffic forecasting and to evaluate their performance in real-world scenarios. To achieve this, state-
of-the-art approaches to traffic forecasting tasks will be explored, in order to find the current trend

1
Introduction 2

of research and explore how to properly deal the main challenges of this problem, such as the
complex relationships between traffic conditions in time, space and external factors. This research
will then be put to use in developing and implementing a machine learning model based on the
current trend of research for road traffic forecasting, considering factors such as weather and time
of day that may affect traffic patterns. Being part of a research project that focuses on smart cities,
the ultimate goal of this dissertation is to be able to use this model to make accurate and reliable
traffic predictions to be incorporated into Intelligent Transportation Systems across different road
networks.

1.3 Contributions
The contributions of this dissertation can be summarized as follows:

• A comprehensive review of the existing literature of traffic forecasting, with a special focus
on techniques based in Graph Neural Networks(GNN).

• Comparison of the performance of GNNs with other commonly used traffic prediction mod-
els in the literature.

• Investigation of the impact of using missing data imputation techniques on traffic prediction
performance of Graph Neural Networks (GNNs).

• Examination of the effectiveness of incorporating external data, such as weather conditions,

in improving the accuracy of traffic predictions.

• Evaluation of the generalizability of the GNN models across different datasets, including
benchmarking datasets and DynamiCITY’s use-case dataset.

• Development of a GNN-based traffic prediction framework that can be applied to different

traffic networks.

Overall, the results of this research contributed to a better understanding of the potential of
GNNs in the context of traffic forecasting.

1.4 Outline
The dissertation begins with an introduction that sets the stage for the research. It provides the
necessary context for understanding the study and outlines the aims and goals that the research
seeks to achieve. The expected contributions of the work are also highlighted, giving an overview
of what can be expected from the document.
The background chapter, Chapter 2, delves into the main areas of study. First, exploring the
concept of Intelligent Transportation Systems (ITS) and their relevance to the research topic. Sec-
ondly, the chapter focuses on the traffic forecasting problem itself. It discusses various traffic
1.4 Outline 3

prediction techniques and categorizes the different types of traffic prediction problems. More-
over, the chapter introduces Graph Neural Networks (GNNs) as a potential solution to the traffic
forecasting problem. The architecture of GNNs is explained, with a particular emphasis on the
attention mechanism.
Chapter 3 reviews the existing literature and research in the field of traffic forecasting. It
provides an overview of various techniques and approaches that have been employed, includ-
ing statistical methods, ensemble learning, hybrid methods, machine learning, and deep learning
techniques. Special attention is given to the application of Graph Neural Networks in traffic fore-
casting. Furthermore, benchmarking datasets used in the field are discussed, and a gap analysis is
presented to identify areas where the current research can make valuable contributions.
The methodological approach chapter, Chapter 4, outlines the methodology followed in the
research. It begins by clearly stating the problem and formalizing it. The chapter then presents
the implementation pipeline, detailing the steps involved in the research process. Validation is
addressed, including the metrics used to evaluate the models and the baselines against which the
models are compared. Additionally, the chapter includes a risk analysis, considering potential
challenges and limitations associated with the research.
Moving on to the implementation in Chapter 5, the focus is on the practical aspects of the
research. It discusses the technologies used in the research and describes the data retrieval and
pre-processing steps. This includes the handling of traffic data, weather data, and event data. The
chapter also explains the process of generating the graph dataset, covering aspects such as adja-
cency matrix generation, traffic speed normalization, and the incorporation of additional features.
Then the base Graph Neural Network architecture is described, including the different layers and
network hyperparameters. Moreover, techniques for handling missing data and incorporating ex-
ternal data, such as weather and event data, are explored.
In the empirical evaluation, Chapter 6, the research findings are presented and analyzed. The
performance and comparisons of the base Graph Neural Network architecture on each dataset
are discussed. The effect of different hyperparameters is evaluated, and the impact of handling
missing data and incorporating weather data is explored. Global comparisons are made to gain
insights into the overall performance of the models. The results are also discussed, providing an
understanding of the empirical findings.
Finally, the dissertation concludes with Chapter 7. The main findings and contributions of the
research are highlighted, and the limitations and potential areas for future research are addressed.
This serves as a reflection on the research journey and offers closure to the document.
Chapter 2

Background

This introductory chapter presents a summary of the research themes explored in this dissertation
so the reader can become familiarized with the fundamental concepts necessary to comprehend
it. Accordingly, this chapter provides an overview of the main topics covered in this dissertation:
ITS, Traffic Prediction and Graph Neural Networks.

2.1 Intelligent Transportation Systems

Intelligent Transportation Systems (ITS) are advanced technologies and systems that are used to
improve the safety, efficiency, and effectiveness of transportation systems (Irawan, Yusuf, & Pri-
hatmanto, 2020). These systems use various technologies, such as computer systems, wireless
communication, and sensors, to gather and analyze transportation data, and then use this data
to control and optimize the performance of the transportation system at hand (Ang, Seng, Ng-
haramike, & Ijemaru, 2022). ITS systems can be applied to a wide range of transportation modes,
including highways, public transit, and active transportation (such as biking and walking). Ex-
amples of ITS include traffic management systems, advanced traveller information systems, and
automated vehicle systems. The goal of ITS is to make transportation safer, more efficient, and
more reliable, while also reducing congestion, emissions, and other negative impacts of transporta-
tion on the environment.

2.2 The Traffic Forecasting Problem

Traffic forecasting is the process of predicting future traffic patterns and conditions. This can in-
clude predicting traffic volumes, travel times, and congestion levels on roads, highways, and other
transportation networks. Traffic forecasting is a critical component of transportation planning and
management, as it enables decision-makers to proactively address potential congestion, improve
safety, and allocate resources effectively. There are various methods used for traffic forecasting,
including traditional statistical models and machine learning (Y. Zhao, Zhang, An, & Liu, 2018).
The goal of traffic forecasting is to provide accurate and reliable predictions, which can be used to

4
2.2 The Traffic Forecasting Problem 5

inform a range of decisions, from planning and designing new transportation systems to optimiz-
ing existing networks.

2.2.1 Traffic Prediction

In the context of artificial intelligence and machine learning, the tasks that try to solve traffic
forecasting problems are referred to as traffic predictions. The terms forecast and prediction can
sometimes be used interchangeably. Still, for the sake of rigour, when referring to the problem
in the transportation domain, the term should be traffic forecasting, whilst the machine learning
approaches should be referred to as traffic prediction techniques.
Road traffic prediction plays a crucial role in traffic management and ITS, and it can be used
to optimize traffic flow, reduce congestion, improve safety, and provide better route guidance and
estimated time of arrival for travellers. With the current advance in technology and the availabil-
ity of real-time traffic data, it’s becoming more accurate, reliable and can be used for different
purposes (Alam, Farid, & Rossetti, 2019).
Traffic prediction is a complex task that poses several challenges. One of the main challenges
is obtaining high-quality data from various sources such as traffic sensors, GPS, and social media,
which can be difficult to obtain, clean and integrate (Kołodziej, Hopmann, Coppa, Grzonka, &
Widłak, 2022). Traffic patterns are influenced by various factors, such as weather, special events,
and human behaviour, making it difficult to model and predict traffic conditions. Handling spatio-
temporal dependencies is also a challenge as traffic conditions are influenced by the interactions
between different parts of the road network, and these interactions can vary over time. With
the increasing amount of data being generated by multiple sources, scalability becomes a major
challenge (Laña, Del Ser, Velez, & Vlahogianni, 2018). Real-time prediction is also challenging as
it requires predictions to be made quickly and accurately. Choosing the right model and evaluating
its performance can also be difficult due to the lack of ground truth data and the dependence of
results on the specific application and data used.

2.2.2 Categorization of Traffic Prediction Problems

Traffic prediction is a broad topic that is categorized into different sub-problems in the literature,
considering the different dimensions of this problem.

• Data Source: Regarding the data source for the problem, it’s possible to have them cat-
egorized into moving, interval and point sources (Angarita-Zapata, Masegosa, & Triguero,
2019) (W. Jiang & Luo, 2022). Interval data sources, such as Automatic Toll Collection
systems, Video cameras, and License Plate Recognition, differ from point detectors in that
they offer a more comprehensive understanding of vehicle movement along a particular seg-
ment of the road. These sources are equipped with sensors placed at two fixed points, which
calculate measures such as travel time between the points (Angarita-Zapata et al., 2019)
(Mori, Mendiburu, Álvarez, & Lozano, 2015). Moving data sources is the newest since
Background 6

it appeared with the use of Global Positioning Systems (GPSs) in vehicles and other de-
vices. These can provide individual traffic data related to vehicles’ trajectories on the roads
and more detailed traffic information (Angarita-Zapata et al., 2019) (Castro, Zhang, & Li,
2012). Point data sources are placed at specific locations on the roads to detect the presence
of nearby vehicles. These sensors, called point data sources, provide information such as
traffic flow ( number of vehicles passing per time unit), occupancy (the percentage of time
that a sensor is detecting a vehicle), and density (the number of vehicles per unit length of
the road). Common types of sensors used in this category include loop detectors, microwave
radars, laser radar sensors, infrared sensors, among others (Lopes, Bento, Huang, Antoniou,
& Ben-Akiva, 2010) (Mori et al., 2015). This is the data source that will be used for this
dissertation.

• Context: The context provides information on transportation environments wherein traffic

predictions can occur and can be defined as urban and freeway (Angarita-Zapata et al.,
2019) (E. Vlahogianni, Golias, & Karlaftis, 2004). The context is important to define the
particularities of the traffic in cause. In urban settings, the traffic speed is lower, and there
are more road intersections and obstacles, wherein highway settings we can expect a higher
speed and fewer road intersections. This dissertation focuses on highway problems.

• Prediction Horizon: The predictions can be short-term, which are usually defined as pre-
dictions for a time interval smaller than 60 minutes or long-term if the prediction window is
bigger than 60 minutes (Irawan et al., 2020) (K. Lee & Rhee, 2022) (Angarita-Zapata et al.,
2019). This work focuses on short-term prediction.

• Prediction Output: Traffic prediction tasks can be classified based on the traffic state to
be predicted. Predictions related to traffic flow, speed, and demand are the most widely
addressed and are typically handled as distinct problems, while other types are generally
grouped together(W. Jiang & Luo, 2022). This work focuses on traffic speed prediction.

• Traffic Representation: The prediction can also be defined in terms of how traffic states
are defined. These are graph/network-based representations or grid-based representations.
Graph-based (R. Jiang et al., 2021). Grid-based traffic prediction refers to the use of a grid-
based data structure, where the area of interest is divided into a regular grid of cells, and
traffic data is collected and analyzed at each cell. This approach assumes that traffic condi-
tions are homogeneous within each cell and that the data collected at each cell can be used
to make predictions about traffic conditions in that cell (D. Chen et al., 2021) (Schörner,
Hubschneider, Härtl, Polley, & Zöllner, 2019) (X. Zhou, Shen, & Huang, 2019). On the
other hand, Graph-based prediction refers to the use of a graph-based data structure, where
the area of interest is represented as a network of interconnected nodes and edges. This
approach is based on the assumption that traffic conditions are influenced by the relation-
ships and interactions between different parts of the network (Cui, Henrickson, Ke, & Wang,
2018) (L. Zhao et al., 2020). This work solely focuses on graph-based representations.
2.3 Graph Neural Networks 7

Figure 2.1: Grid-Based Traffic and Graph-Based Traffic (extracted from (R. Jiang et al., 2021))

Categorizing the various dimensions of the Traffic Prediction problem helps to organize this
broad task into comprehensible subtasks that will be explored in this dissertation, which will focus
on graph-based traffic flow and speed prediction in a freeway context. A diagram that summarizes
the dimensions of Traffic Prediction can be seen in Fig. 2.2

Figure 2.2: Diagram that illustrates the dimension in which traffic prediction problems can be
categorized

2.3 Graph Neural Networks

Deep learning has greatly improved the performance of various machine learning tasks in recent
years, especially where data is represented in a Euclidean space. However, there is an increasing
number of applications where the data is represented as graphs with complex relationships and
interdependencies between objects. This complexity of graph data poses significant challenges
for traditional machine learning algorithms (Z. Wu et al., 2021b). As a result, there has been a
growing interest in developing deep learning approaches for graph data that have culminated in a
new concept within neural networks: Graph Neural Networks (GNNs).
GNNs have become the state-of-the-art approach specially designed to operate on data rep-
resented as graphs which are a powerful representation for many types of data, including social
networks, molecular structures, and in this case, transportation networks (Z. Wu et al., 2021a).
Background 8

Figure 2.3: Comparison between a 2-D Convolution (a) and a Graph Convolution (b) (extracted
from (Z. Wu et al., 2021b))

In GNNs, the nodes in the graph are typically associated with some feature or attribute, and
the edges represent the relationships between the nodes. GNNs learn representations of the nodes
and edges in the graph and make predictions about the graph. They are typically composed of
multiple layers, each of which takes as input the representations of the nodes and edges from the
previous layer and then updates the representations based on the graph structure. The layers of a
GNN use various techniques to learn the graph structure, such as aggregation and pooling, which
allow the network to propagate information through the graph.
One of the key features of GNNs is their ability to handle graph-structured data with variable
sizes and their ability to propagate information through the graph. This allows GNNs to learn from
patterns and relationships in the graph that would be difficult or impossible to learn from using
traditional neural networks that don’t capture this connection, relying on flat data.

2.3.1 General GNN Arquitecture

To better exemplify the architecture of these networks when applied to traffic forecasting problems,
a general model will be presented here. The notation used is explained in table 2.1.
The underlying traffic network can be represented as a graph G = (V, E, A), which can be
weighted or unweighted, directed or undirected, depending on specific tasks. Each node repre-
sents a traffic object, which can be a sensor, a road segment or a road intersection. A is the
adjacency matrix that contains the topology information of the network. Each entry ai j in the ma-
trix represents the proximity between a pair of nodes, which can be binary (only capturing if the
nodes are connected or not) or a numeric value representing e.g. the length of the road connecting
the two nodes (Ye, Zhao, Ye, & Xu, 2022).
Xt is the feature matrix of the G at time t. The features are usually traffic indicators, such
as traffic flow or speed. Given historical indicators of the whole traffic network over past P time
frames, represent as X = [X1 , · · · , Xt , · · · , XP ] aims to predict the future traffic indicators over
the next Q time slices, denoted as Y = [Y1 , · · · , Yt , · · · , YQ ], where Yt represents output graph
with FO features at time t. The problem can be illustrated as in Figure 2.4. This structure can be
used to predict one traffic indicator or various at the same time.
2.3 Graph Neural Networks 9

Table 2.1: Notation used to describe the GNN architecture (adapted from (Ye, Zhao, et al., 2022))

Graph structure elements

G Graph
E Edges of G
V Vertices of G
A = (ai j )N×N ∈ RN×N Adjacency matrix of graph G
Network parameters
N The number of nodes in graph G
FI The number of input features
FH The number of hidden features
FO The number of output features
P The number of past time slices
Q The number of future time slices
Spatio-temporal Features
X ∈ RP × N × FI A series of input graphs composed of N nodes with
FI features over P time slices
Xt ∈ RN×F I An input graph at time t
Xt i ∈ RFI node i in an input graph at time t
Xt, j ∈ RN the jth feature of an input graph at time t
Xt,i j ∈ R the jth feature of node i in an input graph at time t
Y ∈ RP×N×FO A series of output graphs composed of N nodes
with FO features over P time slices
Yt ∈ RN×FO An output graph at time t
Yt i ∈ RFO node i in an output graph at time t

Figure 2.4: Traffic prediction problem formulated as a GNN (extracted from (Ye, Zhao, et al.,
2022))

Accordingly to the literature, GNNs can be categorized into four types, namely, recurrent
GNNs, convolutional GNNs, graph autoencoders, and spatio-temporal GNNs (Z. Wu et al., 2021a),
that will be further explored in Chapter 3.
Because of the spatio-temporal nature of the traffic forecasting problem, the GNNs used in
this field can be categorized as spatio-temporal GNNs. However, specific components of the
other types of GNNs have also been applied to this problem (W. Jiang & Luo, 2022). Spatio-
temporal GNNs are capable of capturing the complex relationships and dependencies between
Background 10

the different road segments, intersections, and the temporal information, which is essential for
accurate traffic forecasting. They can be used to model the temporal dynamics of the traffic on
the graph, by using the graph structure to propagate information between nodes and edges, and
also the temporal dependencies between the different timestamps. This can be achieved by using
temporal convolutions or recurrent layers in the GNN architecture, which allow the network to
learn both spatial and temporal representations of the data (Bui, Cho, & Yi, 2022). A general
representation of these networks can be seen in Figure 2.5.

Figure 2.5: Representation of a spatio-temporal GNN (extracted from (Bui et al., 2022))

[Link] Graph Convolution

One of the most important components of GNNs is the graph convolution. Unlike traditional
convolutions, which are used to process grid-structured data, graph convolutions allow for the
processing of data that is organized as a graph or a network. In graph convolutional networks
(GCNs), each node in the graph is treated as a feature, and the edges of the graph are used to
encode relationships between the nodes (B. Yu, Yin, & Zhu, 2018). The goal of graph convolutions
is to learn a feature representation of the nodes in a graph.
The basic idea behind graph convolutions is to aggregate information from a node’s neighbours
in the graph. This is done by defining a function that takes as input the features of a node and its
neighbours and outputs a new set of features for the node. This function is applied to all the
nodes in the graph in a recurrent manner, allowing for information to flow across the graph and
the features of each node to be updated. The parameters of the graph convolution operation are
learned during the training process (Sergios Karagiannakos, 2021).
The specific way that graph convolutions are implemented can vary, but typically they involve
a combination of matrix multiplication and activation functions. For example, one common ap-
proach involves first representing the graph as a weighted adjacency matrix, where the values in
the matrix represent the strength of the relationships between nodes. This matrix is then used to
2.3 Graph Neural Networks 11

compute a feature representation for each node, which is updated through a series of convolu-
tion operations. These operations can also be performed recursively, allowing for a hierarchy of
representations to be learned, each representing different levels of abstraction.

2.3.2 Attention Mechanism

The attention mechanism in Graph Neural Networks (GNNs) is a computational technique that
allows the model to focus selectively on the most relevant parts of the graph while making predic-
tions (Hamilton, 2020). It allows the model to automatically learn to attend to different parts of
the graph, rather than relying on a pre-defined fixed pooling operation that aggregates information
from the entire graph. The attention mechanism is implemented as a function that takes as input
the node representations and computes a scalar weight for each node. These weights determine
the importance of each node in the graph and are used to compute a weighted sum of the node rep-
resentations, which is used as the graph representation for the next layer (Velickovic et al., 2017).
The attention mechanism allows the model to dynamically allocate computational resources to
different parts of the graph based on their importance for the prediction task, making it well-suited
for complex and dynamic graphs with many inter-dependencies.

Figure 2.6: Schema of attention over one node with respect to its adjacent nodes. (extracted from
(Sanchez-Lengeling et al., 2021))

Note: In the following sections of this dissertation, other specific terms and components re-
lated to GNNs and Machine Learning that were not mentioned in this section will be used, so the
interested reader is redirected to (Ripley, 1996) for the base knowledge of ML and (Hamilton,
2020)(Z. Liu & Zhou, 2020) for a full overview of GNNs and their structure.
Chapter 3

Related Work

After the context provided by Chapter 2, this section aims to summarize the current state of re-
search in road traffic forecasting, providing a review of the latest technical achievements in this
field and highlighting the main challenges that remain for future research in this field and that
guide the work developed in this dissertation.

3.1 Traffic Forecasting Techniques

Road traffic forecasting has been a subject of active research for over 40 years (E. I. Vlahogianni,
Karlaftis, & Golias, 2014). Initially, most approaches relied on traditional statistical methods for
time-series forecasting, such as autoregressive and moving averages models (S. Lee & b. Fambro,
1999) (Smith, Williams, & Keith Oswald, 2002) (Pavlyuk, 2017). However, with the advent of
new technologies and platforms for big data processing and the availability of data from multiple
sources, the focus has shifted to data-driven procedures (W. Jiang & Luo, 2021). Moreover, with
the advent of Machine Learning (ML) and more recently Deep Learning (DL) research has shifted
toward these new approaches for tackling traffic forecasting problems (J. Liu, Wu, Qiao, & Li,
2021). Ensemble techniques and hybrid methods have also been applied.

3.1.1 Statistical Methods

Despite being a more traditional approach to this problem, recent approaches based on time series
forecasting have still been made. In 2019, Cai et al. developed a noise-immune Kalman filter with
maximum correlation entropy for traffic prediction in highways of Amsterdam (Cai et al., 2019).
In 2021 Chuvo et al. experimented with traffic prediction with several time-series forecasting
techniques (Shuvo, Zubair, Purnota, Hossain, & Hossain, 2021). Despite these recent approaches,
the traditional statistical models have certain disadvantages compared to other techniques. They
have a very limited ability to handle external factors, spatio-temporal dependencies and missing
data (Shi & Du, 2022), and other complexities of real-world traffic data. Still, they are able to make
reasonable predictions with small amounts of data (Alsolami, Mehmood, & Albeshri, 2020).

12
3.2 Regression Algorithms 13

3.2 Regression Algorithms

Regression algorithms have also been applied to this problem. Sun et al. propose a local linear
regression model for short-term traffic prediction. The study compares the performance of the
proposed model with previous results of nonparametric approaches using 32-day traffic speed data
collected on a highway in Houston, Texas. The local linear method is found to be better than
the k-NN and kernel methods, as it makes better use of both historical and current data (Sun,
Liu, Xiao, He, & Ran, 2003). More recently, (Alam et al., 2019) explored the use of several
regression models: Linear Regression, SMO Regression, Multilayer Perceptron, M5P model tree
and Random Forest to predict future traffic flow, using real traffic data collected from highways in
Porto, Portugal.

3.2.1 Ensemble Learning

Several articles focus on the application of ensemble learning for traffic prediction. In a 2018
study by Chen et al., the authors compared the performance of different ensemble learning meth-
ods with traditional time series forecasting methods and found that ensemble methods generally
outperformed traditional methods in terms of prediction accuracy (X. Chen, Cai, Liang, & Liu,
2018). Bokaba et al. made a comparative study of the application of several ensemble techniques
to predict traffic congestion in a particular case of a South African highway, improving the pre-
viously used methods in that case (Bokaba, Doorsamy, & Paul, 2022). Yan et al. proposed an
SVR-based ensemble method that was particularly good at dealing with outliers in the dataset and
external data, outperforming simpler SVR methods (Yan, Fu, Qi, Yu, & Ye, 2022).
In general, the literature suggests that ensemble learning can be an effective approach for
improving the accuracy of traffic flow predictions, especially when data from multiple sources is
used, and when it is combined with machine learning models. However, it’s worth noting that
ensemble learning can also be computationally expensive and require more data and resources to
improve performance.

3.2.2 Hybrid Methods

Hybrid methods for traffic forecasting are methods that combine traditional and data-driven ap-
proaches to make predictions. These methods typically involve combining the strengths of differ-
ent models to improve the accuracy of predictions.
(Wang, Li, & Xu, 2017) explored the use of a combination of ARIMA and SVM for traffic
flow prediction by conducting a research study, finding that the hybrid methods resulted in im-
proved prediction accuracy. Developing a hybrid model of ARIMA with a Multilayer Perceptron
(MLP) and a Recurrent Neural Network (RNN) Rajalakshmi et al. (V & S, 2022) achieved better
performance than any of these models achieve on their own. In 2020, Ma et al. (T. Ma, Antoniou,
& Toledo, 2020) also developed a model that combined ARIMA with a neural network achieving
better results than the models individually.
Related Work 14

Overall, hybrid methods have been shown to be effective in improving the accuracy of traffic
flow predictions. By combining the strengths of different models, these methods can capture the
complexity of traffic patterns and the influence of external factors. It’s important to note that
while hybrid methods can be effective in improving the accuracy of traffic flow predictions, they
may also come with higher computational costs and require more data and resources for optimal
performance.

3.2.3 Machine Learning and Deep Learning Techniques

With the increasing amount and diversity of available data and computational power, Machine
learning, particularly Deep Neural Networks (DNNs), have emerged as powerful tools for traffic
prediction. They can automatically learn complex patterns and relationships in extensive amounts
of data and generalise to new, unseen data (Tedjopurnomo, Bao, Zheng, Choudhury, & Qin, 2022).
Among the DNN techniques, Recurrent Neural Networks (RNNs) are the default neural net-
work architecture for handling time-series data. Because of that, they have generally been chosen
over Convolutional Neural Networks (CNNs) for traffic prediction tasks (K. Lee, Eo, Jung, Yoon,
& Rhee, 2021).
Despite the preference for RNNs, it has been demonstrated that CNNs can also be effective for
traffic data because they can better capture its spatial component (K. Lee et al., 2021). This was
shown by Ma et al. in 2017 when they were able to get reasonable prediction in traffic speed by
modelling the spatio-temporal traffic dynamics as images in a two-dimensional time-space matrix
(X. Ma et al., 2017).
As mentioned, research has shown that each of these types of NNs (CNNs and GNNs) are bet-
ter at capturing one of the two dimensions of the spatio-temporal dynamics of traffic data. CNNs
are better at capturing the spatial dimension, while RNNs are better for the temporal dimension.
This made the study of hybrid networks that combine a CNN for the spatial part and an RNN for
the temporal part a natural choice. A representation of such networks can be seen in Figure 3.1.
In 2018, Li et al. constructed a network they called DCRNN (Diffusion Convolutional Recurrent
Neural Network) (Li, Yu, Shahabi, & Liu, 2018), capable of capturing spatial and temporal de-
pendencies using diffusion convolution and the sequence-to-sequence learning framework as well
as scheduled sampling. In 2019, He et al. also created a model based on this concept which they
named STCNN (Spatio-temporal Convolutional Neural Network) (He, Chow, & Zhang, 2019),
combining a CNN with Long Short-term (LSTM) Units, achieving state-of-the-art performance at
the time of publication. These two approaches are still used as baselines for benchmarking more
recent DNN approaches.
Feedforward Neural Networks (FNNs) also play a role in the construction of these hybrid
neural networks for traffic prediction, with their main functions being: aggregating the output of
subnetworks (Yao et al., 2018) or incorporating external data in the network (Y. Wu & Tan, 2016).
Hybrid DNN methods opened the field for the exploration of other architectures that could cap-
ture graph-like structured data, contributing to the emergence of Graph Neural Networks (GNNs).
3.2 Regression Algorithms 15

Figure 3.1: Symplifyed representation of a hybrid CNN-RNN model (extracted from (K. Lee et
al., 2021))

3.2.4 Graph Neural Networks

As mentioned in Section 2.3, Graph Neural Networks (GNNs) are a recent class of deep learning
techniques that are well suited for tasks involving graph data structures, which is the case of traffic
prediction. In this problem, it’s particularly important to capture not only the spatial graph-like
structure of road networks but also the temporal dimension. This notion was first introduced by Yu
et al. in 2018, with their Spatio-Temporal Graph Convolutional Neural Network, which integrates
graph convolutions and gated temporal convolutions using spatio-temporal convolutional blocks
(B. Yu et al., 2018). This model outperformed other state-of-the-art traffic prediction approaches,
leading the way for other researchers to explore these networks and serving as a baseline model
for comparisons in future works.
In 2019, Chen et al. presented another GNN solution with the adoption of a residual recur-
rent architecture that can capture graph-based spatial dependencies and temporal dynamics with
a special emphasis on capturing periodic temporal correlations (C. Chen et al., 2019). The inclu-
sion of external factors in the models was not considered in these previous approaches, so a new
GNN model having these in consideration was proposed by Ge et al.. It uses a graph convolu-
tion layer to capture the spatial correlations, followed by multiple layers of casual convolution to
learn the temporal dynamics. Then, it considers the influence of social factors and road structure
features by using a fusing module (Ge, Li, Liu, & Zhou, 2019). It outperformed approaches such
as STGCN and DCRNN while having a shorter training time and fewer parameters. These re-
sults were attributed to the choice of a simpler architecture but also to the external data inclusion,
hinting towards the fact that these can have a positive impact on predictions. Another approach
was proposed by (Guo, Lin, Feng, Song, & Wan, 2019) based on the attention mechanism. This
model combines the spatial-temporal attention mechanism and the spatial-temporal convolution,
using graph convolutions in the spatial dimension and standard convolutions for the temporal di-
Related Work 16

mension, showing improvement from STGCN and other baseline models. The authors pinpointed
the inclusion of external data as a relevant future work direction.
A problem that can occur when dealing with road network datasets is that the explicit graph
structure does not necessarily reflect the true dependency relation, due to the incomplete connec-
tions in the data, so (Z. Wu, Pan, Long, Jiang, & Zhang, 2019) developed Graph WaveNet, a
graph-based adaption of WaveNet (van den Oord et al., 2016) that focuses on trying to extract an
adaptive adjacency matrix that captures hidden spatial features. Another factor that can influence
traffic predictions is the occurrence of incidents, and for that (Xie et al., 2019) developed a model
for capturing the impact of traffic incidents on traffic flow and speed, showing how this can impact
predictions.
Another proposed architecture is the encoder-decoder, used by (Zheng, Fan, Wang, & Qi,
2019) in GMAN. This architecture features an encoder and a decoder made up of multiple spatio-
temporal attention blocks that model the effect of spatio-temporal factors on traffic conditions.
The encoder processes the input traffic features while the decoder produces the output sequence.
A transform attention layer sits between the encoder and decoder and converts the encoded traffic
features into the decoder’s input. The transformer attention mechanism models the direct relation-
ships between past and future time steps, helping to mitigate the error propagation issue between
prediction time steps. Experimental results demonstrate the superiority of GMAN when compared
to other methods, including Graph WaveNet and STGNN, To try to accommodate global network
dependencies and changes in the graph structure, (Zhang, Chang, Meng, Xiang, & Pan, 2020) pre-
sented an architecture based in Structure Learning Convolution (SLC), which enables extending
the traditional CNN to graph domains and learning the underlying graph structure for traffic pre-
diction. The model has two SLC modules to capture the global and local structures respectively.
Additionally, Pseudo three Dimensional convolution (P3D) networks are combined with SLC to
model the temporal connection. It was tested in 6 different datasets, where it outperformed other
state-of-the-art approaches.
Another significant approach was developed by (L. Zhao et al., 2020), combining GCN and
GRU networks. The GCN is used to capture the topological structure of the graph to obtain the
spatial dependence, and the GRU model is used to capture the dynamic temporal dependence.
The experiments made by the authors showed the robustness of the approach when dealing with
network perturbations. Another technique that uses GRU was developed by (S. Cao, Wu, Zhang,
Li, & Wu, 2022), aiming to address the limitations of conventional graph convolutional networks
GCN in mining global spatial correlations. To achieve this, the authors leverage gated recurrent
units (GRU) and attention to explore local and global temporal correlations simultaneously.
Trying to tackle the missing data problem in traffic prediction, (Zhong, Suo, Jia, Zhang, &
Su, 2021) propose RIHGCN. This method effectively handles missing data through a recurrent
imputation process and a heterogeneous graph structure that captures dynamic spatial correlations
among nodes in the road network. RIHGCN differs from standard GCN models by creating multi-
ple graphs with different edges, which better capture changing spatial correlations over time. The
method integrates data imputation and traffic prediction in a unified framework, optimizing both
3.3 Benchmarking Datasets 17

objectives simultaneously and avoiding accumulated errors. Experiments show that RIHGCN out-
performs existing methods by a significant margin. This work provides an indication that paying
special importance to dealing with missing data can bring a positive effect in the predictions.
Addressing limitations of other methods that only model relationships between node pairs
and node history information, neglecting node properties, the proposed approach by (Hu, Lin,
& Wang, 2022) is a dynamic spatial-temporal graph convolutional network (DSTGCN), which
includes a dynamic graph generation module that adaptively fuses geographical proximity and
spatial heterogeneity information, and a graph convolution cycle module that captures local tem-
poral dependencies. Experiments on two types of traffic prediction tasks show that DSTGCN
outperforms most baseline models, including STGCN, ASTGCN, and GMAN.
In conclusion, Graph Neural Networks have proven to be an effective solution for traffic pre-
diction tasks. Its application has gained significant attention in recent years due to their ability
to capture the complex spatial and temporal dependencies of traffic data. The combination of
Graph Convolutional Networks (GCNs) and Recurrent Neural Networks (RNNs) has been shown
to achieve excellent performance in multi-step traffic flow prediction tasks. Additionally, the in-
troduction of attention mechanisms has further improved prediction results. Furthermore, the
inclusion of external factors, such as the time of day, day of the week, and traffic accidents, in the
prediction process has been shown to have a significant impact on the accuracy of the prediction
results, highlighting the importance of considering the influence of external factors in traffic flow
prediction tasks. Table 3.1 contains a list of all the approaches addressed in this section.

3.3 Benchmarking Datasets

As for most prediction tasks, the data used to extract future predictions plays a significant role
in the traffic forecasting problem. Common data sources used for traffic prediction include loop
detector data, which provides real-time information on traffic volume, speed, and occupancy; GPS
data, which provides information on vehicle location, speed, and trajectory; crowdsourced data,
which is collected from users through mobile apps or social media platforms; weather data, which
provides information on weather conditions; social media data, which provides information on
traffic conditions, travel times, incidents and events reported by users; public transportation data,
which provides information on the location and schedule of public transportation vehicles; and
historical traffic data, which provides information on past traffic conditions, travel times, and in-
cidents. All these sources have their own advantages and limitations, so it is important to evaluate
which data sources are most appropriate for a specific traffic prediction problem (W. Jiang & Luo,
2021). The source this dissertation will focus on is inductive loop detectors, sensors embedded in
the road surface that detect the presence and passage of vehicles. These detectors provide real-
time data on traffic volume, speed, and occupancy, which can be used to estimate traffic conditions,
predict travel times, and identify bottlenecks and incidents.
The benefit of utilizing loop detector sensor data for graph-based modelling is that the ob-
tained traffic information can be directly utilized as node attributes, with minimal pre-processing
Related Work 18

Table 3.1: Compilation of the relevant existing GNN approaches

Prediction
Name Problem Context Tested On
Horizon
STGCN: Spatio-Temporal Graph Convolu- BJER4,
Short-term Speed Highway
tional Network (B. Yu et al., 2018) PeMSD7
MRes-RGNN: Gated Residual Recurrent
Short-term, PeMS-Bay,
Graph Neural Network (C. Chen et al., Speed Highway
Medium-term METR-LA
2019)
GTCN: Temporal Graph Convolutional Net- Short-term, PeMSD4,
Speed Highway
work (Ge et al., 2019) Medium-term PeMSD7
ASTGCN: Attention Based Spatial-
Short-term, PeMsD4,
Temporal Graph Convolutional Network Flow Highway
Medium-term PeMSD8
(Guo et al., 2019)
Graph WaveNet for Deep Spatial-Temporal METR-LA,
Short-term Speed Highway
Graph Modeling (Z. Wu et al., 2019) PeMS-Bay
DIGC-Net: Deep Graph Convolutional Net- Short-term,
work for Incident-driven Traffic Speed Pre- Medium-term, Speed Urban SFO, NYC
diction (Xie et al., 2019) Long-term
Short-term,
GMAN: Graph Multi-Attention Network Urban, PeMs-Bay,
Medium-term, Speed
(Zheng et al., 2019) Highway Xiamen
Long-term
METR-LA,
SLCNN: Spatio-Temporal Graph Structure
Short-term Flow Highway PeMS-Bay,
Learning (Zhang et al., 2020)
PeMs-S
T-GCN: Temporal Graph Convolutional Urban, SZ-Taxi,
Short-term Speed
Network (L. Zhao et al., 2020) Highway Los-loop
ST-UNet: Spatio-Temporal U-Network for METR-LA,
Short-term,
Graph-structured Time Series Modeling Flow Highway PeMS-M,
Medium-term
(T. Yu, Yin, & Zhu, 2019) PeMs-L
RIHGCN: Heterogeneous Spatio-Temporal
Graph Convolution Network for Traffic
Short-term Speed Highway PeMS
Forecasting with Missing Values (Zhong et
al., 2021)
GCRAN: Graph Convolutional Recurrent Short-term, Flow, METR-LA,
Highway
Attention Network (S. Cao et al., 2022) Medium-term Speed PeMS-Bay
PEMS-Bay,
DSTGCN: Dynamic Spatial-Temporal
Flow, Urban, NE-BJ,
Graph Convolutional Network (Hu et al., Short-term
Speed Highway PEMSD4,
2022)
PEMSD8

requirements. However, the sensors are susceptible to hardware failures, resulting in missing or
noisy data, which necessitates the use of pre-processing techniques. Additionally, there are some
drawbacks to using traffic sensor data for graph-based modelling: these sensors can only be in-
3.3 Benchmarking Datasets 19

stalled in a limited number of locations due to factors such as installation cost. As a result, only a
portion of the road network equipped with traffic sensors can be included in the graph, while the
remaining areas would be omitted (W. Jiang & Luo, 2022).
Most GGN-based approaches make their experimental predictions on loop detector datasets,
so having open-source datasets that can be used for benchmarking between different approaches
is important. Consequently, since the advent of GNNs for traffic prediction, some open-source
datasets have become popular and used in different approaches, which are enumerated in Table
3.2 and described as follows.
METR-LA is a dataset that contains traffic speed and volume collected from the highway of
the Los Angeles County road network from March 1st to June 30th, 2012, with 207 loop detectors
in total, with samples aggregated in 5-minute intervals. It is represented in Fig.3.2a.
Caltrans Performance Measurement System (PeMS) (State of California, 2023) is a large-scale
transportation data collection system in California, United States. It is managed by the California
Department of Transportation (Caltrans) and provides real-time and historical traffic data from
over 18,000 vehicle detector stations on the freeway system. The data is collected using various
sensors, including inductive loops, side-fire radar, and magnetometers, and includes information
on the total flow, average speed, and direction of travel for each sensor. Several traffic datasets
have been extracted from PeMS, namely: PeMS-Bay, which can be seen in Fig.3.2b, contains data
from 325 sensors in the Bay Area from January 1st to June 30th, 2017; PeMSD3, which uses
358 sensors in the North Central Area from September 1st to November 30th, 2018; PeMSD4
that contains data from 307 sensors in the San Francisco Bay Area from January 1st to February
28th, 2018; PeMSD7 which uses 883 sensors in the Los Angeles Area from May to June 2012;
PeMSD8: that uses 170 sensors in the San Bernardino Area from July to August 2016.
Seattle Loop (Cui, 2023), represented in Fig.3.2c, is a dataset collected on four connected
freeways (I-5, I-405, I-90, and SR-520) in the Seattle area, from January 1st to 31st, 2015. It
contains the traffic speed data from 323 detectors.
Traffic Speed Guangzhou (X. Chen, Chen, & He, 2018), consists of 214 road segments in
Guangzhou, China (mainly urban expressways and arterials) within two months from August 1 to
September 30, 2016, in 10-minute intervals.

(a) METR-LA (H. Lu et al., 2020) (b) PeMS-BAY (H. Lu et al., 2020) (c) Seatle Loop (Cui et al., 2018)

Figure 3.2: Image representation of benchmarking datasets

Related Work 20

Table 3.2: List of the most popular open-source benchmarking loop-based traffic datasets

Nº of Aggregation Recorded
Name Location Time Period
Sensors Period Metrics
Los Angeles, Speed, March 1st to June
METR-LA 207 5 minutes
CA, USA Volume 26th 2012
San Jose, CA, Speed, January 1st to June
PeMS-BAY 325 5 minutes
USA Volume 30th, 2017
North Central September 1st to
Speed,
PeMSD3 California, 358 5 minutes November 30th,
Volume
USA 2018
San Franciso, Speed, January 1st to Febru-
PeMSD4 307 5 minutes
CA, USA Volume ary 28th, 2018
Los Angeles, Speed,
PeMSD7 883 5 minutes May to June, 2012
CA, USA Volume
San
Speed,
PeMSD8 Bernardino, 170 5 minutes July to August, 2016
Volume
CA, USA
Seattle, CA, Speed, January 1st to 31st,
Seattle Loop 323 5 minutes
USA Volume 2015
Traffic
Guangzhou, August 1 to Septem-
Speed 214 10 minutes Speed
China ber 30, 2016
Guangzhou

3.4 Gap Analysis

In recent years, there has been a growing interest in using graph neural networks (GNNs) for
traffic forecasting. Despite the advances made in this area, there are several research to be done
and improvement point in current research that need to be addressed in order to further improve
the accuracy and applicability of traffic forecasting using GNNs.
One such gap in current research is the handling of missing data, which is a common challenge
in many real-world traffic forecasting applications. This can result from missing observations
due to equipment failures or data errors or from the unavailability of data for some parts of the
network. Some RNN-based models for traffic prediction have explored missing data imputation
(Che, Purushotham, Cho, Sontag, & Liu, 2016) (W. Cao et al., 2018). However, despite the
prevalence of missing data in traffic forecasting, few GNN-based methods have been proposed to
address this issue.
Another gap in the current research is the consideration of external factors (Yin et al., 2022),
such as weather conditions or special events, that can have a significant impact on traffic flow
(Irawan et al., 2020). These external factors can be difficult to incorporate into GNN models,
given the complexity of their interactions with the traffic network. While some efforts have been
made to incorporate external factors into GNN models (Ge et al., 2019) (Ye, Xue, & Jiang, 2022),
more research is needed to fully understand the impact of these factors on traffic flow and to
develop effective methods for incorporating them into traffic prediction models.
3.4 Gap Analysis 21

Interpretability and explainability are also important areas of concern in the use of GNNs for
traffic forecasting (Yuan & Li, 2021) (W. Jiang & Luo, 2022). These models can be complex and
difficult to interpret, making it challenging to understand the reasoning behind their predictions
(Baldassarre & Azizpour, 2019). This lack of interpretability can limit the trust that stakeholders
place in the models and make it difficult to identify areas for improvement. This indicates that
there is a need for more research to develop interpretable and explainable GNN-based models for
traffic forecasting.
Furthermore, as seen in Table 3.1, most GNN approaches focus on short-term traffic predic-
tions, with intervals from 15 to 60 minutes into the future. Long-term prediction has more complex
spatio-temporal dependencies and uncertainty factors (Yin et al., 2022). It is a relevant research
direction to make models that are able to deal with longer predictions.
In addition to these issues, there is also a need for further research on the extraction of addi-
tional information from the output of GNN-based traffic prediction models. For example, (Xie et
al., 2019) have explored the use of GNNs for accident prediction, which is an important applica-
tion of traffic forecasting. However, there is a need for further research to explore other types of
additional information that can be extracted from the output of these models.
Another area that shows a need for further improvement is its general applicability. The differ-
ent approaches analysed in this literature review were tested in a small number of datasets, often
one or two, with little to no focus on exploring how well the models can handle different traffic
networks.
In conclusion, while GNNs have shown promising results in the area of traffic forecasting,
several gaps in current research need to be addressed in order to further improve their accuracy and
applicability. These include the handling of missing data, the consideration of external factors, the
development of interpretable and explainable models, and the extraction of additional information
from the output of these models. By addressing these gaps, it will be possible to further advance
the state of the art in traffic forecasting using GNNs.
Chapter 4

Methodological Approach

This chapter aims to outline the research approach, containing the problem statement and formal
definition, a pipeline of the research process, the validation methods and ends with an analysis of
the risks that can affect the work’s implementation.

4.1 Problem Statement

The main aim of this dissertation is to contribute to the field of traffic forecasting by developing
a more accurate and reliable approach to traffic forecasting using GNNs, that can be incorporated
into Intelligent Transportation Systems.
To achieve this, the dissertation will explore the following research questions:

• What is the suitable design of a Graph Neural Network for traffic prediction?

• How are the models affected by different missing data handling techniques?

• Can external factors, such as weather and nearby social events, be efficiently incorporated
into the GNN models and how do they influence the model’s performance?

• Can the model adapt to different real use cases?

4.2 Problem Formalization

The traffic prediction problem can be viewed as a spatio-temporal problem that involves anticipat-
ing the future traffic state of the entire road network by using historical data of traffic observations
from N sensor stations.
Let the sensor network be defined as a weighted directed graph G = (V, E, A), following the
notation expressed in Table 2.1, where V is the set of N sensor stations, E is a set of roads connect-
ing the roads where the sensors are present and A ∈ RN × N is the adjacency matrix that contains
the distance between sensor stations. Using xti ∈ RF to denote all features (including traffic speed

T
and the added external data) of node i at time t, then Xt = xt1 , xt2 , . . . , xtN ∈ RN×F denotes the

22
4.3 Implementation Pipeline 23

values of all the features of all nodes at time t. X = (X1 , X2 , . . . , Xτ )T ∈ RN×F×τ denotes the
value of all the features of all the nodes over τ time slices. In addition, we set yti = xtf ,i ∈ R to
represent the traffic speed of node i at time t in the future.
So, the problem to be solved becomes that given X , that represents all historical measure-
ments of all the sensors on the network over past τ time slices, being able to predict future traffic

T
states Y = y1 , y2 , . . . , yN  ∈ RN×Tp of all the nodes
 on the whole traffic network over the next
Tp time slices, where y = yτ+1 , yτ+2 , . . . , yτ+Tp ∈ RTp denotes the future traffic state of node i
i i i i

from τ + 1.

4.3 Implementation Pipeline

Figure 4.1 illustrates the methodology followed by this work and summarizes the steps taken to
build the GNN traffic prediction model.
The first task consists in retrieving the necessary data, including traffic data, weather and
events. This is obtained from various sources and formats. So, the next task is to preprocess the
data to make it suitable for the model. This includes cleaning, transforming, and normalizing the
data.
Once the data is preprocessed, the model can be trained so it learns to predict the traffic speed
based on the input features. The model is then tested using the test data to evaluate its performance
using various metrics. Several combinations of parameters will be experimented on, including, for
example, adjusting the model’s learning rate and optimizer.
Based on the testing results of the different parameters, adjustments are made to the training
parameters and new training and consecutively testing takes place.

Figure 4.1: Diagram of the proposed implementation pipeline

4.4 Validation
Having a proper validation stage is an important part of the development since it evaluates the
performance and assesses how well the model is able to generalize to new data and identify any
potential issues.
In this work, the validation is done on three fronts:

• Evaluate the performance on different datasets to evaluate how well the models generalize
and adapt to different circumstances;
Methodological Approach 24

• Compare the performance of the different iterations (base, with missing data handling and
with external factors) of the proposed models to see the impact of these factors in the pre-
dictions.

• Compare the performance with baseline approaches to assess how powerful the proposed
models are against already existing techniques;

4.4.1 Metrics
The metrics that are used in this validation process are Root Mean Squared Error (RMSE), Mean
Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). They are mathematically
defined as:

1
MAE =∑Si=1 ∑Nj=1 xi, j − x̂i, j
N×S (4.1)
1 |xi, j −x̂i, j |
MAPE = N×S ∑Si=1 ∑Nj=1 xi, j (4.2)
q
1 2
RMSE = N×S ∑Si=1 ∑Nj=1 (xi, j − x̂i, j ) (4.3)

Where x̂i, j is the predicted traffic metric and xi, j is the observed traffic metric, with N being
the number of samples in the test set and S the number of sensors. RMSE measures the deviation
of the predictions from the actual values, taking into account the magnitude of the error. A lower
RMSE value indicates that the model’s predictions are closer to the actual values and thus perform
better. MAE measures the average magnitude of the errors in a set of predictions, and its units
are the same as the predicted target, making it easier to interpret the results. However, as it does
not amplify the effect of large errors, it is less sensitive to outliers(Karunasingha, 2022). MAPE
provides a measure of the accuracy of the model’s predictions as a percentage of the actual values,
making it a useful metric for comparing the performance of the model across different datasets.
These metrics were chosen based on their systematic presence in the recent literature as the metrics
for model performance evaluation (Barros, Araujo, & Rossetti, 2015).
Using these metrics, it is possible to quantify the accuracy of the GNN model’s predictions and
compare it to the performance of other methods. The validation process results will provide in-
sights into the strengths and limitations of the explored models, allowing for further improvement
and refinement. The ultimate goal of the validation process is to demonstrate the effectiveness and
feasibility of the explored approaches.
It is important to note that the datasets used contain missing values, and the predictions made
for those cases should not be accounted for in the metrics calculation since they would be altering
the perceived power of prediction on real values. To avoid that, other works used a technique
called masked metrics, in which the missing values are covered by a mask and not used in the
metrics computations. This work also uses this technique.
4.5 Risk Analysis 25

In addition to these metrics, the execution times of the models will also be measured. One
model can produce better results than others, but if it takes a much longer time to train, the dif-
ference in results might not be relevant enough to justify the added computational cost, so it’s
important to also extract some conclusions from this.

4.4.2 Baselines
To make a global comparison between the work carried out and research done in this field, several
baselines will be used. The choice was based on the baselines chosen by other works, with a focus
on time-series methods and the pioneer GNN approaches to the traffic prediction problem. So, the
chosen baselines are the following:

• Historic Average: This approach models the traffic flow as a seasonal process and uses the
weighted average of previous seasons as the prediction (Li et al., 2018). The period used is
one week, and the prediction is based on aggregated data from the previous four weeks. For
example, the prediction for a given Thursday is the average traffic speeds from the last four
Thursdays.

• VAR - Vector Auto-regressive Model: It is a multivariate time series model that captures
the interactions and dependencies between multiple variables. It represents each variable
as a linear combination of its own past values and the past values of other variables. The
VAR model generates predictions by estimating coefficient matrices and a constant term.
The approach was implemented using the statsmodel python package.

• ARIMA: Stands for Autoregressive Integrated Moving Average and is a widely used time
series model. It is a statistical method that combines autoregressive (AR), differencing (I),
and moving average (MA) components to capture the patterns and trends in time series data.
The orders used were (3, 0, 1), and the model was implemented using the statsmodel python
package.

• DCRNN: Is a deep learning model created by (Li et al., 2018) specifically designed for
traffic prediction tasks. It combines concepts from GCNs and RNNs to capture both spatial
and temporal dynamics in traffic data.

• STGCN: Is a deep learning model designed by (B. Yu et al., 2018) to capture both spatial
and temporal dependencies in spatiotemporal data. It combines GCNs with convolutional
and recurrent layers to model spatial and temporal relationships within the data.

4.5 Risk Analysis

Risk analysis is an essential component of any research project as it helps identify potential prob-
lems and assess the impact of these problems on the project’s outcome. A well-conducted risk
Methodological Approach 26

analysis can help minimize potential risks and ensure the research project is completed success-
fully.
One risk that can take place in this area of research is data availability and quality. Traffic
data is crucial for the development and evaluation of the GNN models, but collecting it can be a
challenge, especially if the data is proprietary or if there are privacy concerns (Laña et al., 2018).
Inaccurate or incomplete data can also negatively impact the performance of the GNN models.
To mitigate these risks, this work will utilize benchmarking datasets that have already been used
in many other well-documented approaches, which helps ensure the quality and relevance of said
datasets.
Another risk is the computational resources required for GNN models. GNNs can be com-
putationally intensive and may require significant computational power and time to train (Yin et
al., 2022). This can be a barrier to the practical implementation of GNN models. To address this
risk, it is necessary to have access to high-performance computing resources and accommodate
the appropriate time for training in the work plan. Since one of the goals of this dissertation is to
explore the incorporation of external data in the prediction models, it’s necessary to make correct
use of those so as not to increase the time and computation power needed to train the models in an
unmanageable way.
In addition, there is a risk that the GNN models may not generalize well to new, unseen data
due to overfitting the training data. This can lead to poor performance in real-world applications.
To mitigate this risk, it is necessary to test the performance of the model in different datasets and
adapt it, so it generalizes better.
In conclusion, by considering these potential risks while developing this work, it is possible to
minimize their impact and increase the likelihood of achieving high-quality results.
Chapter 5

Implementation

Following the methodology proposition, this chapter aims to explain the implementation details,
stage by stage. Starting with the choice of frameworks and technologies for the development, then
explaining the data retrieval and preprocessing, followed by the generation of the graph dataset
structure. After this data-focused part, the base GNN architecture generation is explained, fol-
lowed by the additional approaches: handling missing data and adding external data.

5.1 Technologies Used

This work uses Python1 as the primary programming language, and this choice was motivated
by the language’s wide popularity and extensive libraries for data analysis and machine learn-
ing. Python is a widely-used language for machine learning research due to its large and active
community of developers and users, as well as its ease of use and simple syntax. It has a vast
ecosystem of libraries and tools designed for machine learning, which provide access to a wide
range of algorithms, models, and data processing tools. Python is also a flexible and scalable
language that can handle large datasets and complex models, and additionally has interoperability
with other languages and tools, allowing to combine their strengths (Raschka, Patterson, & Nolet,
2020). Overall, Python’s ease of use, its rich ecosystem of libraries, and its flexibility made it a
secure choice for this work.
As for the machine learning framework, PyTorch2 was chosen due to its popularity, flexibility,
and user-friendliness. PyTorch is an open-source machine learning library developed by Meta’s
AI research team, and has quickly gained popularity in the research community due to its dynamic
computational graph and automatic differentiation capabilities. This allows for greater flexibility
in model design and a more efficient training process (O’Connor, 2021).
To help with the creation of the graph-structured machine learning methods, PyTorch Geomet-
ric3 was used. It is a library built on top of PyTorch that provides tools for dealing with learning
on irregularly structured data, such as graphs. It provides a wide range of graph convolutional
1 [Link]
2 [Link]
3 [Link]

27
Implementation 28

layers, graph pooling layers, and graph attention mechanisms that are specifically designed for
graph data.
In traffic forecasting, the temporal dynamics of traffic data are critical, so the use of an-
other library, PyTorch Geometric Temporal, was very helpful. PyTorch Geometric Temporal
(Rozemberczki, 2023) extends PyTorch Geometric to support temporal graph data and provides a
set of temporal graph convolutional layers and temporal pooling layers that can model the tempo-
ral dependencies in traffic data. The library also includes utilities for loading and preprocessing
temporal graph datasets, making it easy to work with real-world spatiotemporal traffic data.
Using these libraries allows a powerful and flexible toolkit for developing and training deep
learning models for traffic forecasting on graph-structured spatiotemporal data. The libraries pro-
vide an efficient, easy-to-use, and well-documented framework for deep learning models, taking
advantage of the latest advances in deep learning.

5.2 Data Retrieval and Pre-processing

As seen in Chapter 3, the effectiveness of traffic prediction models can benefit from integrating
data from multiple sources. In this work, three types of data will be included to predict traffic
flow in major metropolitan areas: loop-sensor-based traffic data, weather data, and events data.
Combining these three sources of data, the aim is to improve the accuracy and robustness of the
traffic prediction model. This section provides a detailed description of each data source, including
its collection methodology, processing, and cleaning.

5.2.1 Traffic Data

A pivotal part of the traffic prediction task is the traffic data used for training and testing the
models. As seen on Section 3.3, the most used data source for this specific problem comes from
loop-detector sensors installed on several points of the road network that measure one or more
metrics for a certain time period.
This work aims at using four different datasets, coming from 3 different sources. METR-
LA and PeMS-BAY are the two most referenced benchmarking datasets used for GNN-based
traffic prediction in the literature, so they will be used to make comparisons easier between the
developed approach and the existing ones. Since PeMS-BAY comes from Caltrans Performance
Measurement System, it was possible to extract a more recent version of the dataset, to test the
use of event data (the availability of events information was only possible for dates starting from
2021). Besides the benchmarking datasets, it was deemed relevant to test the models on another
dataset to further evaluate real-world applicability. So, a dataset with similar characteristics that is
part of DynamiCITY’s research was used. This dataset contains traffic information from "Via de
Cintura Interna" (VCI), the ring road of Porto, and was provided by Infraestruturas de Portugal.
This dataset is not open-source, but since the work is already going to be tested in two other
open-source benchmarking datasets this was not seen as a problem.
5.2 Data Retrieval and Pre-processing 29

On top of the raw sensor traffic data, in order to be able to generate the graph representation of
the traffic, each of the datasets is accompanied by an auxiliary file with the geographic locations
of each sensor.
Table 5.1 lists the main characteristics of the used datasets. The particularities and individual
pre-processing techniques applied to each of the 3 data sources will be explained in the following
subsections.
Table 5.1: Description of each of the different traffic datasets to be used

Number of Period
Name Location Recorded Period
Sensors Length
Los Angeles, CA, March 1st to June
METR-LA 207 4 months
USA 26th 2012
January 1st to June
PeMS-BAY San Jose, CA, USA 325 6 months
30th, 2017
July 1st, 2020 to
PeMS-BAY-2years San Jose, CA, USA 319 2 years
June 30th, 2022
January 1st to June
VCI Porto, Portugal 26 6 months
30th, 2015

[Link] METR-LA and PeMS-BAY

(a) METR-LA Sensor Location (b) PeMS-Bay Sensor Location

Figure 5.1: Sensor location on the benchmarking datasets

Of the benchmarking datasets found in the literature and described in Section 3.3, PeMS-
BAY and METR-LA were the ones chosen to use in this work, because they were the two most
referenced. They are both loop-detector-based datasets from highway road networks, with the
respective sensor location shown in Fig.5.1. Both datasets also follow the same structure, as
illustrated in Fig.5.2: for each time interval of 5 minutes (represented in each row) , there is one
value for each sensor (represented by the columns), which is the harmonic average speed of that
interval in miles per hour (mph).
Implementation 30

Figure 5.2: METR-LA and PeMS-BAY Dataset Structure, where each column represents a sensor
and each row represents a time, with each cell having the harmonic average speed of a sensor in a
timeframe

For PeMS-BAY, since the referenced dataset was collected from Caltrans Performance Mea-
surement System (State of California, 2023), which is a large-scale transportation data collection
system, it was possible to retrieve a new period of data in addition to the referenced period from
January 1st to June 30th, 2017. The new instance of PEMS-BAY data collected dates from the 1st
of July 2020 to the 30th of June 2022, with a duration of 2 years in total. There are two reasons for
this choice: on one side, it comes from the lack of finding relevant event data for older datasets, so
this one can be used to test the incorporation of such factors; on the other side, the bigger temporal
window that this dataset has allows for comparing the results when using longer or shorter dataset
periods for training the models. This new dataset will henceforth be called PeMS-BAY-2years to
distinguish between the two versions. It is relevant to note that the latter version has 319 sen-
sors instead of the original 325 because when retrieving the new data, those sensors’ IDs were no
longer available. Apart from this, these datasets share the same properties.

[Link] Via de Cintura Interna - VCI

This traffic dataset was obtained from 26 loop-detector sensors, whose location can be seen in
Fig.5.3, installed on Via de Cintura Interna (VCI), a highway in Porto, Portugal. VCI is a ring-like
highway that is 21 km long, with two directions separated by a lane separator. The sensors are
positioned transversely in the lane, under all pathways in both directions. They collect information
about the type and number of vehicles passing through every 5 minutes for the specific segment
they are situated in. The historical dataset contains traffic information from 2013, 2014 and 2015,
but in order to have a dataset with a similar length to the benchmark ones, only the last six months
recorded were used.
Unlike the benchmarking datasets described above, the VCI dataset follows a different struc-
ture, with each timestamp having other metrics besides the harmonic average speed value. The
meaning of every field in the original VCI dataset can be seen in Table 5.2. Another relevant fact
is that this dataset contains missing values, which should be addressed and will pose an additional
challenge. For the purpose of this work, it was necessary to change the structure of the dataset to
match METR-LA and PeMS-BAY.
To get the original VCI dataset to match the structure of the benchmarking datasets the fol-
lowing procedure is followed:

• Select the needed time period: records between 01/01/2015 to 30/06/2015.

5.2 Data Retrieval and Pre-processing 31

Figure 5.3: VCI Sensor Location

• The original dataset has predictions in the two traffic directions and only one should be
studied at a time, so only records with Lane_bunndle_direction equal to "C" were kept.

• Select only the needed fields: Equipment_id, Agg_period_start,

Avg_speed_harmonic

• Pivot the dataset table structure to match the structure of PeMS-BAY and METR-LA: Every
value of Equipment_id becomes a separate column that for each Agg_period_start holds the
speed Avg_speed_harmonic value.

• Avg_speed_harmonic values (originally in km/h) are converted to miles per hour.

5.2.2 Weather Data

[Link] Description

To have consistent weather data across all traffic datasets, instead of pre-existing datasets, an
API was used to retrieve historical weather data for each dataset’s locations and time periods.
The chosen API was Weatherbit (Weatherbit, 2023a) for its free availability of high-granularity
historical data for all locations and time periods needed.
For simplicity, only the closest weather station location to the centre of the road network was
picked for each dataset. The smallest frequency of measurements available for historical data is of
15 minutes.

[Link] Pre-processing

To better fit the needs of this work, the weather data needed to go through some pre-processing
steps. First, not all of the dataset fields available4 through the API were needed for the task at
4 Full list of available fields: [Link]
Implementation 32

Table 5.2: Description of each field of the VCI dataset (adapted from (Alam et al., 2017))

Field Data Type Description

Agg_by_lane_bundle_id Integer Unique ID of each record
Agg_id Integer Street ID
Equipment_id Integer Sensor unique ID
Agg_period_start Datetime Beginning time of the interval of measurement
Corresponds to the time between measurements
Agg_period_len_mins Integer
made by the sensors. It is invariably of 5 minutes.
Nr_lanes Integer Number of lanes in the carriageway segment.
The direction of the lane where the data was mea-
Lane_bundle_direction String
sured. Can be “D” (decreasing) or “C” (increasing)
Total_vol Integer Total number of vehicles that went over the sensor
Arithmetic average speed of the vehicles that went
Avg_speed_arithmetic Float
over the sensor
Harmonic average speed of the vehicles that went
Avg_speed_harmonic Float
over the sensor
Average length of the vehicles that went over the
Avg_length Float
sensor
Average space between vehicles that went over the
Avg_spacing Float
sensor
Fraction of the 5-minute interval the sensor was oc-
Occupancy Float
cupied
Light_vehicle_rate Float Ratio of light vehicles that went over the sensor
Number of vehicles of class A that went over the
Vol_class_A Integer
sensor
Number of vehicles of class B that went over the
Vol_class_B Integer
sensor
Number of vehicles of class C that went over the
Vol_class_C Integer
sensor
Number of vehicles of class D that went over the
Vol_class_D Integer
sensor
Number of unidentified vehicles that went over the
Vol_class_0 Integer
sensor
Axle_class_vol String Number of vehicles per vehicle axis number.
 5.2 Data Retrieval and Pre-processing 33

hand, so they were not included, such fields were: UV index, solar radiance, atmospheric and
sea level pressure, humidity and solar angles. Also, despite being relevant for traffic prediction,
the snow-rate field was also removed, given that for the cities concerned in the traffic datasets,
its value was always zero. The ’pod’ field was converted from string to int, ’n’ became 0 and ’d’
became ’1’.
The final number of fields became 6 out of the original 24 made available by the API. The
fields that were kept can be found in Table 5.3.

Table 5.3: Description of the fields from the weather API that were used (adapted from
(Weatherbit, 2023b) )

Field Data Type Units Description

temp Float Celsius Temperature
wind_spd Integer m/s Wind speed
clouds Integer Percentage Cloud Coverage
pod String - Part of the day. Can be “d” (day) or “n” (night)
visibility Integer Km How far it is possible to see
precip_rate Float mm/h Precipitation Rate

As previously mentioned, the smaller granularity available for the historical data was of 15-
minute intervals, while the traffic datasets have 5-minute intervals. So there was the need to match
these intervals so the data could be used together. Given the bigger importance of the traffic data
over the weather data, it did not make sense to reduce the granularity of the traffic data to 15-
minute aggregates, because some information would be lost by doing that. Instead, the weather
dataset was modified to have 2 more records between each 15-minute interval by doing linear
interpolation of the available records. It is known that this introduces some inaccuracy in the data,
but its effects were considered less damaging than reducing the traffic dataset granularity.

[Link] Automation of weather dataset generation

To allow other researchers to follow the same procedure and generate weather datasets to use
with loop-sensor data for traffic prediction tasks, a command line tool that can generate weather
datasets with the characteristics as described above for any location and time period was created.
It retrieves the data from the Weatherbit API, processes it and saves it to a CSV file.
To use the tool, there’s the need to input several parameters, such as a valid Weatherbit API
key, the latitude and longitude of the location for which to generate the weather data, as well as
the start and end dates for the data and finally the name for the output CSV file. The usage is as
follows:

1 python weather_tool.py <API_KEY> <LATITUDE> <LONGITUDE> <BEGIN_DATE>

2 <END_DATE> <OUTPUT_FILE>

Listing 5.1: Weather tool usage

Implementation 34

So, running:

weather_tool.py abc123 37.7749 -122.4194 2022-01-01 2022-01-31

my_weather_data.csv

would build weather data for San Francisco, CA from January 1, 2022 to January 31, 2022, using
the Weatherbit API key ‘abc123‘, and save it to a file named ‘my_weather_data.csv‘.

5.2.3 Event Data

[Link] Description

The events dataset was more difficult to find than the previous ones since there is less availability
of open-source quality data regarding this matter, and there was the need to find events occurring
in specific timeframes in specific locations. There was only one relevant data source found that
had the needed information, which was the PredictHQ API (PredictHQ, 2023a). The PredictHQ
API aggregates and enriches event data from various sources to provide a comprehensive and up-
to-date view of events happening around the world. It includes information such as event type,
location, time, duration, and attendance. Unfortunately, it is only possible to retrieve the historic
events data for free if they are less than a year in the past, counting from the moment of the API
call, so it created the need for having a more recent traffic dataset, as previously explained.
The choice of events to retrieve from the API was based on the following criteria:

• The location of the event needs to be within a 20 km radius of the road network;

• The event’s occurrence needs to be between the recorded time period of the traffic dataset;

• The attendance should be of at least 1000 people, this was deemed a reasonable threshold
between small gatherings and events that have the potential to significantly affect the way
people usually move in the city;

• The event belongs to one of the following categories: conferences, expos, concerts, festi-
vals, performing arts, sports, community, public holidays, school holidays, severe weather
and disasters. The excluded categories were non-attendance-based events, such as daylight
savings, which were considered to not have an impact on mobility.

[Link] Pre-processing

In terms of pre-processing, the data provided by the API did not need much work to be ready for
use in this work, so the pre-processing procedure was simple. One of the required steps consisted
in removing the variables that would not be relevant to this use case. The event records directly
retrieved contained 32 different fields (PredictHQ, 2023b), with information such as if the event
was private or public, if it had been rescheduled or postponed when it was announced, and many
others that don’t need to be included, so they were removed. In addition, the location of the event
was originally expressed as a complex datatype with several sub-fields, so it was changed in order
 5.2 Data Retrieval and Pre-processing 35

to only have two separate fields of longitude and latitude. Table 5.4 shows the fields and respective
descriptions of the processed dataset. After these pre-processing steps, the final events dataset for
PeMS-BAY-2years is comprised of 394 events, with the category distribution as shown in Figure
5.4.
Table 5.4: Description of each field of the Events dataset

Field Data Type Description

category String Type of event (concert, holiday, sports, etc...)
start Datetime Timestamp of the start of the event
end Datetime Timestamp of the end of the event
lat Float Latitude of the event’s location
long Float Longitude of the event’s location
attendance Integer Event’s attendance

Figure 5.4: Category distribution of the events retrieved for PeMS-BAY-2years

[Link] Automation of events dataset generation

To allow other researchers to follow the same procedure and generate events datasets to use with
loop-sensor data for traffic prediction tasks, a command line tool that can generate events datasets
with the characteristics as described above for any location and time period, as long as they are
from less than one year in the past from the moment the API call is being made, was created. It
retrieves the data from the PredictHQ API, processes it and saves it to a CSV file.
To use the tool, there’s the need to input several parameters, such as valid PredictHQ API
client ID, API secret and access token, the latitude and longitude of the ’epicentre’ for which to
generate the data, as well as the start and end dates for the data in and finally the name for the
output CSV file. The usage is as follows:

1 python event_search.py <CLIENT_ID> <API_SECRET> <ACCESS_TOKEN> <LATITUDE>

2 <LONGITUDE> <RADIUS> <BEGIN_DATE> <END_DATE> <OUTPUT_FILE>
Implementation 36

Listing 5.2: Events tool usage

So, for example, running:

python events_tool.py 1234 secret token 40.7128 -74.0060 10

,→ 2023-03-01 2023-03-31 nyc_events.csv

would build a dataset with events in New York City within a radius of 10 kilometres between
March 1, 2023 and March 31, 2023, and export the results to a file called nyc_events.csv

5.3 Graph Dataset Generation

After the retrieval and pre-processing of data, the next step is constructing the input to feed the
GNNs from the original data.
Since this work will be tested on several different datasets, one of the goals was being able to
create a uniform way of assembling the GNN input data from any given dataset, both to ease the
modular development of this work and to further increase its future usability for other use cases,
i.e. other road networks or different recorded time periods. For this, it was important to determine
the information needed to be able to generate the input data, which was ultimately narrowed down
to this set of files :

• A csv file that represents the speed values per timestamp in every sensor, with the structure
shown in Fig.5.2, henceforth called traffic data;

• A csv file with information about the coordinates of the sensors, with the structure as shown
in Figure 5.5.

Figure 5.5: Structure of the file with the sensors’ location

Given these files, it is possible to generate the input data for the GNN models for any given
road network and time period by following the procedures shown in Fig.5.6, which will be further
described in the next subsections.
5.3 Graph Dataset Generation 37

Figure 5.6: Description of the graph dataset generation

It is important to note that this data generation phase is adapted to add the additional features
of handling missing data and external data incorporation, which will be described in Section 5.5
and Section 5.6, respectively.

5.3.1 Adjacency Matrix Generation

[Link] Calculating Road Distance Between Sensors

Having the information from the sensors’ location is not enough to extract all the spatial infor-
mation of the road network since the road structure and direction of traffic flow also play a part.
Having an adjacency matrix that represents the distances between the sensors is needed. For
PeMS-BAY and METR-LA datasets, the distance between sensors was made available by the de-
velopers of the DCRNN approach in their GitHub repository (Li, Mensi, & Yu, 2023). However,
for the VCI dataset, it was not available, so there was a need to create it. It may seem trivial to ob-
tain these distances by calculating them from the distance between the coordinates, but this would
be a very rough estimation of the true adjacency of the road distance (i.e. the distance that a car
would need to travel from an origin to a destination) between the sensors. So, an alternative was
found by using Openrouteservice5 , which is a free-of-charge and open-source API, which pro-
vides road distances between points. Like this, it was possible to create a way to build adjacency
matrices for this particular VCI dataset and, in general, for other datasets that might be used in the
future.

[Link] Calculating Edge Weights

Having the road distance between sensors, it is now possible to calculate the weight of the graph’s
edges. Instead of using the raw distance values between sensors to fill the adjacency matrix, this
work follows the technique used in DCRNN(Li et al., 2018) and STGCN(B. Yu et al., 2018), in
which the weighted adjacency matrix W is filled using the Thresholded Gaussian Kernel function
(Shuman, Narang, Frossard, Ortega, & Vandergheynst, 2013):
5 [Link]
Implementation 38


exp − [dist(vi ,v2 j )]2
 
2σ
if dist(vi , v j ) ≤ κ
Wvi ,v j = (5.1)
0 otherwise

where Wvi ,v j is the edge weight between sensor vi and sensor v j , dist (vi , v j ) represents the road
network distance from vi to v j , σ is the standard deviation of distances and κ is a threshold for
sparsity, meaning that sensors with a weight lower than the threshold are set to 0. This is especially
helpful for bigger networks where sensors very far from each other have almost no interaction,
and therefore their connection can be ignored by the network model to speed up calculations. The
threshold k was set to 0.1 as it was done by the approaches mentioned above.

5.3.2 Traffic Speed Normalization

Normalizing data for neural networks is an important step to take since it helps optimization algo-
rithms by balancing the scales of input features, preventing imbalanced gradients. It accelerates
the convergence process by enabling quicker identification of the optimal solution and enhances
generalization by reducing sensitivity to variations in absolute magnitude and focusing on under-
lying patterns within the data (Bhanja & Das, 2019).
The normalization method chosen was Z-score normalization since it is more robust with
outliers than min-max normalization and doesn’t restrict the data to a given range.

5.3.3 Additional Features

On top of the speed sensor readings, adding additional temporal features that could help capture the
temporal dependencies between the data was deemed important. So, considering the base traffic
dataset as an array with shape [T, S] , where T is the number of timestamps and S the number of
sensors, in this step, another dimension is added so for each sensor and timestamp instead of only
having the traffic speed information, there can be other features, making the shape [T, S, F], where
F is the number of features (traffic speed included).
The explored features were:

• Time in day: every five-minute interval of the day gets a different value between 0 and 1
sequentially, so the same time on different days shares this feature to try to capture daily
patterns.

• Day in week: all records that are from the same day of the week get the same value to try to
capture weekly patterns.

• Hour in day: all records that are from the same hour of the day get the same value.

• Is weekend: all records that belong to weekends get 1 and others get 0

The goal here is to train the models with different combinations of these features to see if
they help the results. Here only temporal features extractable from the traffic dataset itself were
explained. Data from other sources are explored in Section 5.6.
5.3 Graph Dataset Generation 39

5.3.4 Building Inputs and Targets

The models explored throughout this work are sequence-to-sequence models: for a given times-
tamp, the observed feature values of that timestamp plus an additional number of immediately
preceding values are taken as historical features to predict x future speed values. From now
on, the number of historical features and the number of future prediction horizons will be called
’num_historical_ f eatures’ and ’num_prediction_horizons’, respectively.
Having this in mind, when creating the input and target tensors for the model, they will have
the following shapes [T, S, F, num_historical_ f eatures] and [T, S, num_prediction_horizons].
The input tensor has four dimensions, meaning that the model takes as input for each sensor in
each timestamp ’num_historical_ f eatures’ of historical features preceding the given timestamp.
The output tensor has three dimensions meaning that the model for each sensor in each timestamp
generates ’num_prediction_horizons’ number of predictions into the future, each with 5 minutes
between them. The tensor only has the speed information and not the additional features present
in the input tensor since these are not being predicted and are only used to generate the speed
predictions.
After having the dataset in this final form, it can be split for the train, validation and test.

5.3.5 Train, Validation and Test Splitting

In terms of splitting the dataset, in this case, since Neural Networks are used, and they need to
have their hyperparameters tuned during training, on top of a train and test sets, a validation set is
also needed.
The splits are done in the time dimension, i.e. each subset contains a portion of the timestamps
of the dataset but the entire set of sensors.
In most cases, this split should be done by randomly selecting which samples go into each
subset. However, since the data in question has a temporal aspect, this choice needs to be re-
considered. Time-series data has a chronological order and, therefore, some form of temporal
dependencies. This is different from many other types of data, where the observations are usually
considered to be independent of each other. By randomly splitting time-series data, some "future"
data would likely be in the training set and some "past" data in the test set. This violates the
temporal order and dependency of the data, and it results in "data leakage", where the model has
access to future information it wouldn’t have at prediction time (Jensen, 2019). This can lead to
overly optimistic performance estimates during training and validation stages, which might not
hold when the model is applied to real-world, future data.
When splitting the data chronologically (i.e., the training set consists of all data up to a certain
time point, followed by validation and then the test set), the temporal order of the data is main-
tained, which is a more realistic approach and represents how the model will be used in practice:
trained on past data and used to make predictions for future data. Furthermore, chronological
splitting allows the model to capture possible trends, seasonality, and other temporal dynamics in
the training data, which might be relevant for future predictions.
Implementation 40

5.3.6 Dataloader
After having the data split, the last step is to use PyTorch’s ’DataLoader’ method to divide inputs
and targets into batches of a specified size, allowing for more efficient processing. It makes an it-
erable object from the batches of inputs and targets, allowing iteration over the batches for training
and test loops.
The data in this form, together with the adjacency matrix generated, are what is fed the model
to obtain future predictions.
5.4 Base Graph Neural Network Architecture 41

5.4 Base Graph Neural Network Architecture

In this study, the capabilities of PyTorch Geometric and PyTorch Geometric Temporal libraries
will be used to adapt GNN architectures for spatiotemporal traffic prediction. These libraries pro-
vide a comprehensive set of building blocks and tools that facilitate the design and implementation
of GNNs. This part of the work aims to discover which architecture achieves superior performance
across various datasets.
The process begins with the exploration of candidate architectures with a set of base hyperpa-
rameters configuration, on all datasets. This serves as a primary evaluation of the performance of
the different models, allowing the identification of the most promising architectures. Once the two
top-performing architectures have been identified, the next step is to evaluate their performance
with other hyperparameters.
Upon selecting the optimal GNN architecture and the set of optimal hyperparameters, the aim
is to try to enhance the performance further by incorporating external data sources and implement-
ing strategies to handle missing data.

5.4.1 Layers
The models explored in this work follow the structure illustrated in Fig.5.7. The data goes through
N Spatio-temporal blocks made up of a spatial graph mechanism followed by a temporal mech-
anism. Several mechanisms were tried, as enumerated in Table 5.5. The number of blocks is
a parameter that was explored in the empirical experiments. After the spatio-temporal blocks,
the data goes through a fully connected layer, followed by ReLU activation, to result in the final
network output.

Table 5.5: Enumeration of the temporal and spatial mechanisms that were explored

Spatial Graph Mechanisms

GATConv Graph Attention Convolution
GCNConv Graph Convolution
ChebConv Chebyshev Spectral Graph Convolution
Temporal Mechanisms
GRU Gated Recurrent Unit
CNN Convolutional Neural Network
GRU + Attention GRU + Temporal Attention
CNN + Attention CNN + Temporal Attention

[Link] Spatial Graph Mechanism

This part of the network is responsible for spatial dependency modelling, capturing the relationship
between neighbouring sensors in the network.
The mechanisms that were explored are the following:
Implementation 42

Figure 5.7: Structure of the GNN layers

• GAT - Graph Attention Network: GAT leverages attention mechanisms to model the rela-
tionships between nodes in a graph. GAT was introduced in (Velickovic et al., 2017) to cap-
ture the importance of neighbouring nodes when aggregating information during message-
passing iterations. In GAT, each node in the graph is associated with a learnable attention
mechanism. Attention allows the model to assign different weights to the neighbouring
nodes based on their importance in the context of the target node. This mechanism works
with multi-head attention, where k different mechanisms run in parallel, originating k inde-
pendent outputs that are then concatenated and linearly transformed. This parameter k is set
to 3 in this work, as was done in the original paper.

• GCN - Graph Convolution Network: The graph convolution operation was introduced
in (Kipf & Welling, 2017). It updates a node’s features by aggregating its neighbouring
nodes’ features. This process is repeated for several layers, each time considering a larger
neighbourhood, which allows the network to learn more global features of the graph.

• ChebConv - Chebyshev Spectral Graph Convolution: This graph operation was intro-
duced in (Defferrard, Bresson, & Vandergheynst, 2016), and it is a variation of GCN de-
signed to approximate spectral graph convolutions by using Chebyshev polynomials. It
operates in the graphs’ spectral domain, leveraging the Laplacian matrix’s eigenvalues and
eigenvectors. By truncating the expansion of the Chebyshev polynomials, it efficiently com-
putes convolution on graph signals without the need for costly eigenvalue decomposition.
This operator allows for the propagation of information and capturing of spatial dependen-
cies in graph data.
5.4 Base Graph Neural Network Architecture 43

[Link] Temporal Mechanism

This part of the network is responsible for temporal dependency modelling, capturing the influence
of past sensor readings with more recent ones.
The mechanisms that were explored are the following:

• Gated Recurrent Unit: The most commonly used neural network model for processing
sequence data (i.e. temporal data) are RNNs. However, traditional RNNs have long-term
prediction limitations due to gradient disappearance and explosion. To address these prob-
lems, variants of RNNs such as the LSTM and GRU have been developed. As LSTM has a
more complex structure and longer training time, while the GRU model has a simpler struc-
ture and faster training ability, the GRU model is chosen to obtain temporal dependence
from traffic data in some works, such as (Li et al., 2018) (L. Zhao et al., 2020). The GRU
model uses a hidden state from the previous time step and the current traffic information
as inputs to predict the traffic status at the current time step. This allows the model to cap-
ture the current traffic information while retaining the changing trends of historical traffic
information and capturing temporal dependence.

• CNN: Despite RNNs being more commonly used in temporal data, CNNs are also used to
model this part of the data in works such as (B. Yu et al., 2018) due to their faster training
times, simpler structures, and lack of dependency on previous steps. Inspired by (Gehring,
Auli, Grangier, Yarats, & Dauphin, 2017) the authors propose a method that employs en-
tire convolutional structures on the time axis to capture the dynamic behaviours of traffic
flows. This design allows for parallel and controllable training procedures. The temporal
convolutional layer comprises a one-dimensional causal convolution with a width-Kt kernel,
followed by Gated Linear Units (GLU) for non-linearity. This process essentially maps the
input (which could be seen as a multi-channel time series) to a single output element via a
convolutional kernel.

• Gated Recurrent Unit + Attention: This mechanism was used in (Zhu, Song, Zhao, &
Li, 2020), has an extension of their previous work (L. Zhao et al., 2020), that after passing
through GRU, the hidden states are inputted into an attention model to determine the context
vector that covers the global traffic variation information in time.

• CNN + Attention: This mechanism was introduced in (Guo et al., 2019), and it consists of
combining a CNN with an attention mechanism to capture the temporal dependencies.

5.4.2 Network Hyperparameters

One of the challenges of this work is to identify the optimal combination of parameters that yield
the best performance for the specific network architectures that are being explored. By system-
atically investigating the impact of different parameter settings, the aim is to understand their
influence on the overall performance and generalization capabilities of the GNN models.
Implementation 44

The importance of selecting appropriate network parameters cannot be overstated, as it directly

affects the model’s capacity to learn and adapt to the underlying data patterns. A well-tuned GNN
model can significantly improve accuracy and robustness while reducing the risk of overfitting and
ensuring efficient training.
This section describes the parameters that were explored in this work. It is important to note
that the parameter exploration needs to be kept to a reasonably minimal size not to create a com-
bination too large to be experimented on.

• Learning Rate: The learning rate is a crucial hyperparameter determining the step size
during optimization. A suitable learning rate can significantly improve the model’s conver-
gence speed and overall performance. However, selecting an appropriate learning rate can
be challenging, as it often depends on the problem domain, data distribution, and model ar-
chitecture. While a larger learning rate leads to a premature, sub-optimal solution, a smaller
one will converge the optimisation process slowly (K. Zhou, Liu, Duan, & Hu, 2022). Al-
ready existing approaches used either 0.01 or 0.001 as a base learning rate. In this work,
the base learning rate was set to 0.01 on the first tests done, but on top of those, some other
experiments with other learning rates, including 0.001, were also carried out.

• Training Epochs: The number of training epochs is a key hyperparameter that determines
how many times the model will update its parameters based on the training data. Balancing
the number of epochs is essential to avoid underfitting, where the model fails to capture the
underlying patterns in the data, and overfitting, where the model becomes overly specialized
to the training data and performs poorly on unseen instances. Recent GNN studies (Li et
al., 2018)(B. Yu et al., 2018)(Guo et al., 2019)(L. Zhao et al., 2020) indicate the use of 100
epochs for training their GNN models, so the same is applied in this work.
An important detail to note is that the model that gets saved from training to be applied to
the test data and evaluated is not the model that results from the last epoch. but rather the
model from the epoch that got the lowest loss in the validation data. This means that in the
training loop the current validation loss is always checked against the best until that point ,
and if it is even lower, the model gets saved.

• Loss Function: Most GNN works use MSE (Li et al., 2018) (B. Yu et al., 2018) (T. Yu et
al., 2019) loss, so this was also applied in this work. It is important to note that since these
works don’t do any missing data pre-processing when calculating the loss function, they use
a variation of the original definitions, where the loss is only calculated based on non-missing
values by applying a mask to null values. These are often called masked loss functions.

• Optimizer: The optimizer is the algorithm that adjusts the network’s weights and biases
to minimize the loss function. It does this by computing the gradients of the loss function
according to the parameters and updating them iteratively. The optimizer’s goal is to find
the best parameter values that reduce the loss, improving the network’s predictions(Lau,
2017). Previous works with GNN mostly use the Adam optimizer. In addition to Adam,
5.5 Handling Missing Data 45

other optimizers, AdamW and AdaGrad, will be used in the experiments to see if they can
produce better results.

• Train, Validation and Test proportions: This work used 80% of data for training, 10% for
validation and 20% for testing.

• Number of input features and output targets: As mentioned in Subsection5.3.4, when

building the input data, n historic features are taken and the output contains x predictions.
These numbers are configurable and so different combinations of input features and output
targets were explored, but the default value is 12 inputs and 12 outputs, meaning one hour’s
worth of data is taken as input and the model generates 12 predictions, for 5 minutes after
the origin timestamp, 10 minutes after, 15 minutes after and so on until one hour.

• Number of spatio-temporal blocks: Other techniques that use a similar consecutive spatio-
temporal blocks approach show better results with 2 blocks (Guo et al., 2019) (L. Zhao et
al., 2020), so this was the number used in the experiments of this work.

5.5 Handling Missing Data

As previously explained, the existing techniques regarding traffic speed prediction with GNNs do
not attempt to handle the missing data present in the datasets they use but instead mask them when
calculating loss and evaluation metrics. So in this work, it was explored if dealing with the missing
data can positively impact the predictions made by the models. But first, it is important to explore
the characteristics and quantity of the missing data present in the datasets.

5.5.1 Missing data Occurrence and Characteristics

After analysing the datasets, it was possible to observe different types of occurrences of missing
data. In the case of VCI, some sensors showed no readings for long periods of time, sometimes
for more than a month, as shown by plotting the entire set of sensor readings in Fig.5.8a, in which
the first 1.5 months are not present. This can mean that the sensor in question was broken and it
took that amount of time for it to be replaced or fixed.
On top of these blocks of missing data, there also happen sparse missing data, when in the
middle of valid sensor values are missing values or zeros. These can happen by sensor failure but
can have a different reason: the sensors record the average speed of vehicles passing through them
in the given time period, so if for a period of 5 minutes, no vehicle passes through, then the sensor
will register 0 as the speed value for that period, which is also not a valid reading. An example of
this can be seen in Fig.5.8b, plotting only one day of a sensor’s readings, and there are several 0
values during the late night hours.
In terms of the number of missing values, the four datasets used in this work have distinct
proportions, as shown in Table 5.6. PeMS-BAY and PeMS-BAY-2years are significantly less
Implementation 46

affected by missing values than the others. METR-LA, despite not having blocks of null values in
its sensors, has a significant amount of sparse zeros, more than in the VCI dataset.

(a) Block of missing data

(b) Sparse missing data in the form of zeros

Figure 5.8: Types of occurring missing data in the sensors (VCI dataset)

Table 5.6: Presence of missing values in the dataset

Ratio of null values Ratio of zeros Total ratio of missing values

METR-LA 0% 8.11% 8.11%
PeMS-BAY 0% 0.03% 0.03%
VCI 4.09% 0.24% 4.33%
PeMS-BAY-2years 0.01% 0% 0.01%

5.5.2 Missing Data Handling Techniques

After identifying the missing data, the next step is to find appropriate ways of dealing with them,
keeping in mind the special characteristics of the data at hand.
The problem as it is represented in this work, regards the traffic data as a spatio-temporal
dataset, with the readings of one sensor having a correlation with other sensors, so removing the
missing data is not an option in this case. Instead, data imputation techniques will be explored.
For the choice of these techniques, since most of them work sensor by sensor (not considering the
5.5 Handling Missing Data 47

network correlations) and each sensor individually can be considered a time series of traffic speed
records, the techniques chosen to apply should be ones used in time series data.

[Link] Mean

This is a simple and commonly used technique for handling missing data. The missing values are
replaced with the mean value of the sensor readings. Used with time series data it has limitations
since it does not account for temporal dependencies between sensor readings. The plotted values
of the imputation in a sample of data from the VCI dataset can be seen in Fig.5.9a for a block of
missing values and Fig.5.9b for sparse missing data.

(a) Block imputation

(b) Sparse imputation

Figure 5.9: Comparison between original data and after mean imputation in VCI

[Link] Last Observation Carried Forward and Next Observation Carried Backward

Last Observation Carried Forward (LOCF) and Next Observation Carried Backward (NOCB) are
two distinct imputation techniques used in time series data. LOCF uses the last observed value
to fill in missing values that occur after it and NOCB uses the next observations to fill in missing
values that occur before it. In this case, they were used together, LOCF before NOCB to fill the
missing values since there can be values missing at the beginning and end of the dataset that would
be left missing if only one of the techniques were used. The choice of performing LOCF before
NOCB comes from past values having influence over future values, so there was the desire to keep
that relationship whenever possible. The plotted values of this imputation in a sample of data from
Implementation 48

the VCI dataset can be seen in Fig.5.10a for a block of missing values and Fig.5.10b for sparse
missing data.

(a) Block imputation

(b) Sparse imputation

Figure 5.10: Comparison between original data and after LOCF + NOCB imputation in VCI

[Link] Linear Interpolation

It assumes a linear relationship between the observed values before and after the missing value
and estimates the missing value based on this assumption. The plotted values of this imputation in
a sample of data from the VCI dataset can be seen in Fig.5.11a for a block of missing values and
Fig.5.11b for sparse missing data.

[Link] MICE - Multiple Imputation with Chained Equations

MICE starts by calculating the column-wise mean for any columns containing missing values and
replaces those missing values with the mean. It then proceeds to perform a series of chained
regression models to impute each missing value iteratively (Ismiguzel, 2022). Similar to standard
regression, MICE utilizes a feature matrix and a target variable for training. The plotted values of
this imputation in a sample of data from the VCI dataset can be seen in Fig.5.12a for a block of
missing values and Fig.5.12b for sparse missing data.
5.5 Handling Missing Data 49

(a) Block imputation

(b) Sparse imputation

Figure 5.11: Comparison between original data and after interpolation imputation in VCI

5.5.3 Comparing the techniques

By looking at the different techniques plotted values compared with the rest of the data, it is
possible to determine how well they might perform. Still, before using them for the GNN model,
it is helpful to get more information on their power to impute values as close to as they would be
if they were real. Since for actual missing values, it isn’t possible to know what the sensor reading
would be it is not possible to see how the techniques perform against each other empirically.
To overcome that, a subset of the VCI dataset was created, containing only sensors with no
missing values (which ended up being only four). Then, for that subset, missing values were
artificially generated with the following configurations:

• Sparse 4%: Only sparse missing values randomly selected at 4% missing rate for the subset;

• Sparse 8%: Only sparse missing values randomly selected at 8% missing rate for the subset;

• Block I: Only block missing values, randomly selecting two sensors, removing one month’s
worth of data from one at a random location and from the other removing two weeks also at
a random location;

• Block II: Only block missing values, randomly selecting two sensors, removing two month’s
worth of data from one at a random location and from the other removing one month also at
a random location;
Implementation 50

(a) Block imputation

(b) Sparse imputation

Figure 5.12: Comparison between original data and after MICE imputation in VCI

• Sparse + Block I: Sparse 4% + Block I used together

• Sparse + Block II: Sparse 4% + Block II used together

Then, the performance of each imputation technique in the different missing data configura-
tions was evaluated using MAE and RMSE metrics. To remove possible bias related to the random
choice of the missing data, the metrics were generated by averaging over five runs with different
randomly generated missing data. The results can be seen in Table 5.7, and they show that for
sparse missing data using interpolation produces the best results, but for blocks of missing data,
the MICE approach produces smaller errors. In response to this, a hybrid between these two ap-
proaches that uses MICE for blocks and interpolation for sparse missing values was created in it
produced the best results for when the both types of missing data exist.

Table 5.7: Empirical evaluation results of the different imputation techniques

Sparse (4%) Sparse (8%) Block I Block II Sparse + Block I Sparse + Block II
MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE
Mean 0.368 2.739 0.739 3.886 0.307 1.794 0.623 2.643 0.719 3.670 1.300 5.155
LOCF +NOCB 0.154 1,228 0.309 1.728 0.366 2.074 0.786 3.057 0.586 2.903 1.059 3.937
Interpolation 0.128 0.960 0.257 1.360 0.297 1.880 0.703 2.957 0.575 2.670 0.961 3.705
MICE 0.275 2.129 0.565 3.060 0.294 1.585 0.616 2.359 0.605 3.019 1.210 4.431
MICE+Interpolation 0.128 0.960 0.257 1.360 0.294 1.585 0.616 2.359 0.561 2.402 0.955 3.154
5.6 External Data Incorporation 51

5.5.4 Inclusion in the Data Processing Pipeline

This additional data processing step is added to the existing data processing pipeline described
in Section 5.3, before the traffic data is normalized, so the processing steps become as shown in
Figure 5.13.

Figure 5.13: Data processing pipeline with missing data imputation

5.6 External Data Incorporation

After the base GNN architecture and experimenting with missing data handling techniques, the
addition of data from external sources, which were described in Section 5.2, was conducted.

5.6.1 Weather Data

One of the data retrieved to complement the traffic data was the weather conditions in the zone
where the road network in question is located, at every timestamp recorded. This choice came
from the general notion that on days when there are bad weather conditions, such as heavy rain,
low visibility and/or low temperatures there seems to be more congestion in roads. This could
have many possible reasons, such as more people driving by car instead of public transportation
or walking so as to not be subject to the weather, making the network more congested. This idea
is supported by the data used in this work: using VCI as an example, the traffic conditions of
2 days were chosen to be analysed. One day in which the weather conditions were very good
(high visibility, no rain, little to no clouds and no significant winds) and one where the weather
conditions were very bad (low visibility, heavy rain throughout most part of the day, and significant
winds). Both days were the same day of the week (both Sundays) so as to not have weakly patterns
affect the analysis as much. The comparison can be seen for three different sensors in Fig.5.14.
The chosen sensors are located in distanced places of the network, so as to show that the weather
impact is shown across the network.
The weather data was added as an additional feature, as was described for the additional tempo-
ral features in Subsection 5.3.3. Several different combinations of such weather condition features
were used, which are the following:

• precip_rate + visibility
Implementation 52

(a) Comparison in the traffic of sensor ’11’

(b) Comparison in the traffic of sensor ’12’

(c) Comparison in the traffic of sensor ’23’

Figure 5.14: Comparison between the traffic in a bad weather day and a good weather day

• precip_rate + visibility + wind_spd

• precip_rate + wind_spd

• precip_rate + wind_spd + visibility + temp

• all features: precip_rate + wind_spd + visibility + temp + pod + clouds

This list is not by any means the full exhaustion of all the possible combinations of features,
but given the constraint in terms of training time available, it was deemed a reasonable set of
possibilities concerning the most promising features. The results of this inclusion can be seen in
Section 6.3.
5.6 External Data Incorporation 53

5.6.2 Event Data

The other kind of data retrieved to complement the traffic data was the occurrence of high-
attendance events in the area close to the road network in question. This choice came from the
fact that such events by having high amounts of people gathering in a specific location at a specific
time will directly lead to an influx of transportation needs to that location at that time, changing
the usual behaviour of the network’s flow.
This effect can be seen in the recorded traffic data, for example, Fig.5.16a shows the evolu-
tion of the traffic speed through 24 hours in one of VCI’s sensors, represented by number 5 in
Fig.5.15 during two different days. One of them, designated ’match day’ represents Sunday 20th
of September, 2015, in which, according to Futebol Clube do Porto’s official agenda6 there was
a match between them and Sport Lisboa e Benfica at 19:15 held at Estádio do Dragão, which
is located very close to the VCI road, in particular to the sensor in observation. The other day
represented is the Sunday of the week before. In the figure, it is noticeable the decrease in speed
observed in the late afternoon, particularly in the times preceding the beginning of the match and
then in the evening around the time the match ended, both not present in the ’normal’ day. How-
ever, this effect is not transversal to the whole network, in Fig.5.16b by plotting the same days
but in sensor ’25’, far away and in the opposite direction of Estádio do Dragão, it is possible to
see that the speeds don’t differ as they do in sensor ’5’. This shows how this interaction between
events and traffic is more complex than, for example, the weather since it has different effects in
different parts of the network and also at different times.
It is important to note that a single case doesn’t prove that all events will significantly impact
the network’s traffic, but it indicates that it is worth exploring this dependency.

Figure 5.15: VCI Sensor Location and traffic flow

6 [Link]
Implementation 54

(a) Comparison in the traffic of sensor ’5’

(b) Comparison in the traffic of sensor ’25’

Figure 5.16: Comparison between the traffic during a match day in Estádio do Dragão and a
regular day

Unfortunately, given time restrictions related to the computational time and power required to
run the models, it was not possible to carry out this exploration. This means that the PeMS-BAY-
2years dataset and the corresponding retrieved event dataset were not used. However, given that
the datasets were retrieved and processed, it made sense to keep the mentions to them since they
can be used in future research on the topic.
Chapter 6

Empirical Evaluation

In this chapter, the empirical experiences carried out will be shown and discussed. Given their
big computational time requirements, different experiments ran on different machines at the same
time to economize time. It was relevant to register the training time of the different models, but
this became challenging when using different machines for different experiences since they had
slightly different performances. The compromise that was deemed adequate for this work was,
since there isn’t a big discrepancy in the time of different epochs of the same model, to run 10
epochs of each model all on the same machine and register the average of those epochs. During
those runs, there was an effort not to have other processes running on the machine so as to have
execution time comparisons as reliable as possible. The machines had the same hardware speci-
fications, which were the following: Intel core i7-8700K @3.70GHz processor, 32 GB of RAM
and an Nvidia GeForce GTX 1080 graphics card running on Windows 10. The source code was
developed in Python 3.10, with PyTorch 2.0, running with CUDA 11.8. Ensuring similar results
will only be possible by having these hardware specifications and library versions. To simplify
the identification of the different models explored, they were named in this chapter following the
components of their network structure, as described in Table 6.1.

Table 6.1: Identification of the names used for the different GNN arquitectures

Name Network
GAT-GRU GAT as spatial mechanism and GRU as temporal mechanism
GAT-CNN GAT as spatial mechanism and CNN as temporal mechanism
GAT-GRUAtt GAT as spatial mechanism and GRU+Attention as temporal mechanism
GAT-CNNAtt GAT as spatial mechanism and CNN+Attention as temporal mechanism
GCN-GRU GCN as spatial mechanism and CCN as temporal mechanism
GCN-CNN GCN as spatial mechanism and CNN as temporal mechanism
GCN-GRUAtt GCN as spatial mechanism and GRU+Attention as temporal mechanism
GCN-CNNAtt GCN as spatial mechanism and CNN+Attention as temporal mechanism
Cheb-GRU ChebConv as spatial mechanism and CCN as temporal mechanism
Cheb-CNN ChebConv as spatial mechanism and CNN as temporal mechanism
Cheb-GRUAtt ChebConv as spatial mechanism and GRU+Attention as temporal mechanism
Cheb-CNNAtt ChebConv as spatial mechanism and CNN+Attention as temporal mechanism

55
Empirical Evaluation 56

6.1 Base Graph Neural Network Architecture

As explained in the methodology, here, the different GNN Architecture explored were compared
in the three datasets, and the two best-performing ones on a base configuration of hyperparameters
were chosen to be further explored with different hyperparameter settings.

6.1.1 Model Comparisons in Each Dataset

For this experiment, the settings were the following:

• Learning rate: 0.01

• Number of training epochs: 100

• Optimizer: Adam

• Loss function: Mean Squared Error

• Historical Features: 12

• Prediction Horizons: 12

• Number of Spatio-temporal blocks: 2

• Time features: using only the ’Time in day’

Table 6.2 shows the performance of all the explored models in regard to the PeMS-BAY
dataset. It shows the values of MAE, MAPE and RMSE in 3 prediction horizons: 15, 30 and
60 minutes, as well as the average time each epoch takes to train in seconds. It is important to note
that the models generate predictions for 12 horizons, but only these three are shown since having
them all displayed would make the results very dense and hard to understand. This principle will
be followed throughout the tables of this Chapter.

Table 6.2: Comparison of the different models on the PeMS-BAY dataset

15 min 30 min 60 min Training time
Models
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE per epoch (s)
GAT-GRU 3.706 8.009 6.491 4.551 9.921 6.957 5.348 10.672 7.848 108.345
GAT-CNN 3.763 8.015 6.471 4.698 9.915 6.955 5.245 10.933 7.756 116.324
GAT-GRUAtt 3.579 8.217 6.167 3.869 9.215 6.500 4.647 10.755 7.410 111.876
GAT-CNNAtt 3.409 7.988 6.064 3.799 8.904 6.409 4.543 10.543 7.220 120.320
GCN-GRU 4.359 9.002 6.500 4.892 10.543 8.205 5.994 11.418 8.894 101.456
GCN-CNN 4.466 9.107 6.545 4.952 10.567 8.189 6.082 11.608 8.955 107.980
GCN-GRUAtt 3.722 8.126 6.596 4.619 9.978 7.142 5.411 10.890 7.912 105.876
GCN-CNNAtt 3.804 8.199 6.708 4.685 10.001 7.150 5.587 11.103 7.992 110.435
Cheb-GRU 3.724 8.107 6.455 4.599 9.901 6.944 5.407 10.683 7.798 129.435
Cheb-CNN 3.748 8.122 6.399 4.651 9.915 7.074 5.290 10.901 7.795 121.003
Cheb-GRUAtt 1.869 3.920 3.341 2.406 5.428 4.495 3.077 7.463 5.736 131.120
Cheb-CNNAtt 3.420 8.057 6.103 3.886 8.998 6.497 4.612 10.602 7.295 123.327
6.1 Base Graph Neural Network Architecture 57

One thing that all models have in common results-wise is that the predictions get progressively
worse for bigger horizons, which is in line with what is expected since the further away in time
we are, the more difficult it is to make good predictions.
Comparing the results of the same temporal mechanism used with different spatial mecha-
nisms, it is noticeable that GCN produces the worst results out of the three mechanisms, which
is in line with the fact that this technique is less advanced and was then improved with Cheby-
shev polynomials to become GhebConv. Comparing ChebConv with GAT is not as clear. The
results show similar performances across almost all temporal mechanisms. There is the exception
of Cheb-GRUAtt, which significantly outperforms the other ChebConv-based models.
In terms of comparing the results of the same spatial mechanism used with different tempo-
ral mechanisms, the temporal mechanisms that have temporal attention (GRUAtt and CNNatt)
significantly outperform the ones without attention, so it can be said that having the attention
helps to extract the most relevant time dependencies. In terms of choosing between GRUAtt and
CNNAtt, GRUAtt tends to have better results but not by a big difference, with the exception of
GAT-CNNAtt, that is better than GAT-GRUAtt.
Now looking into other datasets, Table 6.3 shows the performance of all the explored models
in the METR-LA dataset.
Table 6.3: Comparison of the different models on the METR-LA dataset

15 min 30 min 60 min Avg. time

Models
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
GAT-GRU 5.420 15.704 8.524 5.755 17.079 9.108 6.265 19.633 10.111 53.498
GAT-CNN 5.446 15.728 8.567 5.806 17.112 9.143 6.234 19.678 10.099 50.349
GAT-GRUAtt 4.978 15.626 8.851 5.342 16.648 9.219 5.899 18.606 9.974 58.457
GAT-CNNAtt 4.877 15.501 8.713 5.190 16.547 9.157 5.785 18.457 9.869 54.630
GCN-GRU 5.890 15.954 8.785 6.102 17.789 9.324 6.654 19.996 10.302 48.768
GCN-CNN 5.921 16.023 8.798 6.132 17.819 9.335 6.678 19.952 10.343 46.674
GCN-GRUAtt 5.551 15.799 8.620 5.865 17.240 9.193 6.370 19.769 10.206 51.348
GCN-CNNAtt 5.557 15.823 8.652 5.898 17.255 9.204 6.362 19.787 10.199 49.540
Cheb-GRU 5.472 15.734 8.562 5.769 17.102 9.112 6.276 19.648 10.102 81.234
Cheb-CNN 5.493 15.789 8.581 5.797 17.123 9.145 6.291 19.667 10.109 78.965
Cheb-GRUAtt 2.995 8.201 5.621 3.622 10.748 6.876 4.586 14.770 8.456 83.439
Cheb-CNNAtt 4.983 15.637 8.865 5.358 16.632 9.207 5.909 18.651 9.921 80.735

Again, the predictions get progressively worse for bigger horizons, as in PeMS-BAY. The
performance of the models in METR-LA is worse than in PeMS-BAY, in general across all models.
The performance of the temporal and spatial mechanisms follows the tendencies seen in PeMS-
BAY.
Table 6.4 shows the performance of all the explored models in the VCI dataset.
Again, the performance of the different mechanisms follows the trend of the previous two
datasets.
Comparing the results across the three datasets, some conclusions can be extracted:

• The different models show the same relative performance against each other (i.e. if a model
performs better than some other model in one of the datasets it also performs better in the
Empirical Evaluation 58

Table 6.4: Comparison of the different models on the VCI dataset

15 min 30 min 60 min Avg. time

Models
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
GAT-GRU 5.129 9.318 8.458 5.8298 10.612 9.001 5.778 11.213 9.907 72.789
GAT-CNN 5.138 9.337 8.476 5.8265 10.643 9.024 5.785 11.227 9.9019 64.120
GAT-GRUAtt 4.620 8.615 7.745 5.123 10.612 8.798 5.643 11.110 9.899 75.675
GAT-CNNAtt 4.418 7.642 7.353 4.807 8.692 8.265 5.412 10.254 9.495 67.212
GCN-GRU 5.965 10.132 8.945 6.043 11.537 9.788 6.655 11.997 10.337 60.245
GCN-CNN 5.973 10.097 8.959 6.051 11.548 9.736 6.678 12.006 10.341 57.127
GCN-GRUAtt 5.551 9.799 8.620 5.865 11.240 9.224 6.437 11.769 10.064 62.900
GCN-CNNAtt 5.512 9.650 8.589 5.920 11.320 9.193 6.370 11.710 10.178 59.976
Cheb-GRU 4.993 9.255 8.487 5.103 10.855 8.555 5.625 11.899 9.737 119.765
Cheb-CNN 5.009 9.267 8.496 5.148 10.855 8.543 5.678 11.850 9.702 116.953
Cheb-GRUAtt 4.017 6.750 6.677 4.601 8.453 7.892 5.451 10.668 9.450 122.710
Cheb-CNNAtt 4.345 6.925 6.754 4.601 8.591 8.214 5.667 10.826 9.633 118.678

other datasets) on the different datasets, which makes it possible to consider that they are
capable of being applied to different use cases.

• The results are overall better in PeMS-BAY, followed by VCI and then METR-LA. This
could have to do with the ratios of missing value occurrence in each dataset, given that
PeMS-BAY has a very small amount of missing values, then VCI sits in the middle, and
METR-LA has the most.

• In terms of training time, the models have the same relative performance across all datasets,
but the same model in different datasets takes different times. This makes sense since the
datasets have different dimensions, both in sensor number and duration of the recorded
period. Models take the most time to run on PeMS-BAY, followed by VCI and then METR-
LA.

• Models using GRU as the temporal mechanism tend to be slower than the ones using CNN,
which goes in line with the theory behind the two architectures.

• The temporal attention mechanism also minimally slows down training.

• Models with ChebConv as their spatial mechanism are slower than those with GAT or GCN.
PeMS-BAY is the largest dataset in the number of sensors (it has 325) and has the same
recorded period as VCI (6 months), but VCI has only 26 sensors. METR-LA has 206
sensors and 4 months of recorded period. This gives the indication that the recorded period
length has more impact than the number of sensors. Otherwise, the models would run faster
in VCI than in METR-LA.

From now on, as they showed better performance in all datasets and for all prediction times,
the Cheb-GRUAtt and GAT-CNNAtt models will be the ones in which the following experiments
will be performed.
6.1 Base Graph Neural Network Architecture 59

6.1.2 Effect of Hyperparameters

After the top-performing models with the initial hyperparameters configuration were discovered,
being Cheb-GRUAtt and GAT-CNNAtt, these models were tested with different hyperparameter
configurations to try to improve their performance further.

[Link] Effect of the Optimizer

For this experiment, the settings, apart from the optimizer, were the following:

• Learning rate: 0.01

• Number of training epochs: 100

• Loss function: Mean Squared Error

• Historical Features: 12

• Prediction Horizons: 12

• Number of Spatio-temporal blocks: 2

• Time features: using only the ’Time in day’

Tables 6.5, 6.6, and 6.7 contain the results of the different optimizers in the PeMS-BAY,
METR-LA and VCI datasets, respectively, for the Cheb-GRUAtt model. In regards to this model,
in METR-LA, the optimizer with the best results is Adam across all metrics and prediction hori-
zons. In PeMS-BAY, this only verifies on 30 and 60-minute horizons, and for 15 Adagrad gets
better results. In the case of VCI, the results are not as clear: for a 15-minute horizon, Adagrad is
better, but for 30 and 60-minute horizons the results are very similar across all optimizers.
In terms of the time needed for training, the difference between the optimizers is minimal.
Given that, for two of the datasets (METR-LA and PeMS-BAY), Adam is better, and in VCI
the difference isn’t very significant, the following experiences with this model will be performed
using the Adam optimizer.

Table 6.5: Comparison of the different optimizers used with the Cheb-GRUAtt on the PeMS-BAY
dataset

15 min 30 min 60 min Avg. time

Optimizers
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
Adam 1.869 3.920 3.341 2.406 5.428 4.495 3.077 7.463 5.736 131.120
AdamW 1.887 3.886 3.383 2.428 5.390 4.527 3.186 7.640 5.785 132.450
Adagrad 1.714 3.835 3.353 2.306 5.482 4.605 3.143 7.960 6.056 132.970

Now, for the GAT-CNNAtt model, Tables 6.8, 6.9, and 6.10 contain the results of the different
optimizers in the PeMS-BAY, METR-LA and VCI datasets, respectively. Here the predominance
of better results with Adam that was seen on the previous model does not verify. In METR-LA
Empirical Evaluation 60

Table 6.6: Comparison of the different optimizers used with the Cheb-GRUAtt on the METR-LA
dataset

15 min 30 min 60 min Avg. time

Optimizers
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
Adam 2.995 8.201 5.621 3.622 10.748 6.876 4.586 14.770 8.456 83.439
AdamW 4.390 17.542 8.373 4.743 18.421 8.880 5.325 19.900 9.697 83.248
Adagrad 4.442 17.670 8.489 4.848 18.603 8.905 5.554 20.242 9.707 83.952

Table 6.7: Comparison of the different optimizers used with the Cheb-GRUAtt on the VCI dataset

15 min 30 min 60 min Avg. time

Optimizers
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
Adam 4.017 6.750 6.677 4.601 8.453 7.892 5.451 10.668 9.450 122.710
AdamW 3.954 6.905 6.653 4.545 8.565 7.853 5.402 10.663 9.229 121.437
Adagrad 3.764 6.578 6.605 4.397 8.433 7.906 5.270 10.783 9.476 124.789

Table 6.8: Comparison of the different optimizers used with the GAT-CNNAtt on the PeMS-BAY
dataset

15 min 30 min 60 min Avg. time

Optimizers
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
Adam 3.409 7.988 6.064 3.799 8.904 6.409 4.543 10.543 7.220 120.32
AdamW 3.089 6.734 5.816 3.199 6.994 5.925 3.340 7.513 6.151 120.45
Adagrad 2.715 6.293 5.085 2.834 6.531 5.246 3.1525 7.276 5.735 120.81

Table 6.9: Comparison of the different optimizers used with the GAT-CNNAtt on the METR-LA
dataset

15 min 30 min 60 min Avg. time

Optimizers
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
Adam 4.877 15.501 8.713 5.190 16.547 9.157 5.785 18.457 9.869 54.630
AdamW 4.882 15.929 8.657 5.057 16.497 9.059 5.459 17.808 9.883 56.604
Adagrad 3.840 11.725 6.957 4.606 12.764 7.496 4.670 14.983 8.477 55.490

Table 6.10: Comparison of the different optimizers used with the GAT-CNNAtt on the VCI dataset

15 min 30 min 60 min Avg. time

Optimizers
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
Adam 4.418 7.642 7.019 4.807 8.365 8.265 5.413 10.254 9.495 67.212
AdamW 4.501 7.738 7.313 4.907 8.796 8.318 5.545 10.435 9.325 63.962
Adagrad 4.610 8.346 7.750 4.947 9.141 8.269 5.552 10.586 9.201 64.365

and PeMS-BAY, Adagrad is better across all metrics and horizons. This does not hold for VCI,
where Adam performs better.
Comparing the results of the two models, it is not possible to conclude that one particular
optimizer is better than the other for this problem, given that the two models work better with
different optimizers. One conclusion that is possible to extract is that the optimizers don’t have a
significant influence on the training time. It is also possible to observe that the results in VCI tend
to work differently than for the other datasets. This dataset is a bit more distinct from the others,
having a smaller network of sensors that are spaced more apart, which could explain the different
6.1 Base Graph Neural Network Architecture 61

behaviour, but that isn’t enough information to conclude that.

[Link] Effect of the Learning Rate

For this experiment, the settings, apart from the learning rate, were the following:

• Optimizer: Adam (for the Cheb-GRUAtt model) / Adagrad (for the GAT-CNNAtt model)

• Loss function: Mean Squared Error

• Historical Features: 12

• Prediction Horizons: 12

• Number of Spatio-temporal blocks: 2

• Time features: using only the ’Time in day’

The goal here was to test learning rates smaller and bigger than 0.01 (the value used on the
previous tests). Given the expected behaviour of experimenting with learning rates, starting with
a bigger learning rate (that moves too fast and skips the optimal weights), the results would get
progressively better until a ’sweet spot’ is reached, where the optimal weights can be reached
within the given epochs, and after that, the results would get worse again, since more epochs would
be needed to converge. Following this principle, the procedure followed was to try progressively
lower learning rates until the results started deteriorating.
Table 6.11 has the results of training the Cheb-GRUAtt model with different learning rates
over 100 epochs on PeMS-BAY. It is possible to see that the learning rates 0.001 and 0.0005 have
the best results, with 0.001 beating 0.0005 in the 15-minute horizon but 0.0005 is better for the
bigger horizons.

Table 6.11: Results of different learning rates with 100 epochs for Cheb-GRUAtt on PeMS-BAY
Learning Nr of 15 min 30 min 60 min Avg. time
Rate Epochs MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
0.020 1.842 3.963 3.403 2.502 5.643 4.630 3.268 7.915 5.972 131.674
0.015 1.825 3.964 3.373 2.392 5.490 4.545 3.195 7.698 5.857 130.322
0.010 1.869 3.920 3.341 2.406 5.428 4.495 3.077 7.463 5.736 131.120
0.005 1.789 3.793 3.295 2.345 5.369 4.467 3.069 7.420 5.687 132.002
100
0.0025 1.709 3.695 3.248 2.274 5.286 4.420 3.004 7.328 5.638 131.107
0.001 1.673 3.659 3.236 2.275 5.256 4.419 3.029 7.319 5.660 130.992
0.0005 1.678 3.673 3.241 2.266 5.231 4.425 2.979 7.214 5.647 130.717
0.00025 1.687 3.667 3.247 2.270 5.242 4.434 2.985 7.285 5.669 131.203

For METR-LA, the results are in Table 6.12 and don’t show such a clear picture, with 0.005
being better.
Table 6.13 contains the results in the VCI dataset, where it is possible to see that the best
learning rate across all horizons and metrics was the smallest one tested, 0.00025, with results
getting better and better the lower the learning rate is. This could indicate that an even smaller
Empirical Evaluation 62

Table 6.12: Results of different learning rates with 100 epochs for Cheb-GRUAtt on METR-LA
Learning Nr of 15 min 30 min 60 min Avg. time
Rate Epochs MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
0.020 2.908 7.907 5.480 3.548 10.456 6.721 4.483 14.190 8.241 83.135
0.015 2.916 7.911 6.732 3.547 10.396 6.732 4.467 14.112 8.248 82.854
0.010 2.995 8.201 5.621 3.622 10.748 6.876 4.586 14.770 8.456 82.450
0.005 3.139 7.907 5.493 3.532 10.350 6.728 4.408 14.028 8.261 83.123
100
0.0025 2.900 7.922 5.001 3.533 10.438 6.750 4.477 14.206 8.258 83.637
0.001 2.902 7.944 5.496 3.544 10.434 6.740 4.493 14.136 8.237 82.500
0.0005 2.877 7.832 5.500 3.535 10.349 6.748 4.523 14.159 8.266 83.439
0.00025 2.895 7.825 5.508 3.593 10.457 6.781 4.646 14.504 8.334 83.127

value could still provide better results, but since in the other datasets, this small learning rate got
worse results than some bigger ones, the tests stopped at 0.00025.
By observing the results on the three datasets, it is possible to say that they follow the same
tendencies but not exactly the same results, not being able to pinpoint one learning rate that sys-
tematically is better in all three datasets. Since the goal of this work is to try to generalize the
process to fit different datasets, it was decided that 0.0005, which was the best performing on
PeMS-BAY but also worked well with the two others, would now be used in all the remaining
tests.

Table 6.13: Results of different learning rates with 100 epochs for Cheb-GRUAtt on VCI
Learning Nr of 15 min 30 min 60 min Avg. time
Rate Epochs MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
0.020 4.094 6.968 6.842 4.687 8.773 8.091 5.572 11.215 9.696 120.065
0.015 4.107 7.174 6.853 4.686 8.870 8.082 5.534 11.062 9.624 119.645
0.010 4.017 6.750 6.677 4.601 8.453 7.892 5.451 10.668 9.450 122.710
0.005 3.893 6.560 6.574 4.471 8.227 7.754 5.309 10.439 9.223 120.602
100
0.0025 3.845 6.399 6.493 4.452 8.191 7.703 5.274 10.432 9.167 120.749
0.001 3.835 6.554 6.493 4.425 8.237 7.697 5.229 10.353 9.148 121.689
0.0005 3.800 6.452 6.458 4.397 8.180 7.658 5.210 10.399 9.151 120.941
0.00025 3.791 6.394 6.445 4.375 8.076 7.621 5.181 10.285 9.077 118.870

Now concerning the other model, GAT-CNNAtt, the result in PeMS-BAY, METR-LA and VCI
datasets can be seen, respectively, in Tables 6.14, 6.15 and 6.16. For PeMS-BAY, the best learning
rate of the ones tested is 0.05, for METR-LA it is 0.015 and for VCI the results are not as clear
between 0.020 and 0.015. As it happenned with the other model, different datasets have better
results with different learning rate, but 0.015 was the one chosen to continue the next tests in this
model.
In terms of the effect the learning rate has on the training time, it does not give a noticeable
influence.

[Link] Effect of the additional time features

For this experiment, the settings, apart from the learning rate, were the following:

• Number of training epochs: 100

6.1 Base Graph Neural Network Architecture 63

Table 6.14: Results of different learning rates and with 100 epochs for GAT-CNNAtt on PeMS-
BAY
Learning Nr of 15 min 30 min 60 min Avg. time
Rate Epochs MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
0.020 2.642 6.364 5.071 2.741 6.654 5.247 3.094 7.173 5.620 121.193
0.015 2.751 6.567 5.211 2.849 6.766 5.359 3.117 7.393 5.754 121.505
0.010 100 2.715 6.293 5.085 2.834 6.531 5.246 3.1525 7.276 5.735 120.810
0.005 2.602 6.172 4.933 2.727 6.447 5.123 3.009 7.170 5.594 121.225
0.0025 2.921 6.803 5.256 3.020 7.041 5.401 3.327 7.797 5.841 121.505

Table 6.15: Results of different learning rates and with 100 epochs for GAT-CNNAtt on METR-
LA
Learning Nr of 15 min 30 min 60 min Avg. time
Rate Epochs MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
0.020 4.150 13.000 7.582 4.337 13.789 7.950 4.762 15.578 8.691 54.899
0.015 3.683 10.872 6.687 3.968 12.068 7.301 4.530 14.435 8.344 55.184
0.010 100 3.840 11.725 6.957 4.606 12.764 7.496 4.670 14.983 8.477 55.490
0.005 4.318 13.233 7.733 4.572 14.380 8.281 5.063 16.213 9.130 55.062
0.0025 4.541 14.630 8.012 4.781 15.567 8.421 5.193 17.069 9.088 54.851

Table 6.16: Results of different learning rates with 100 epochs for GAT-CNNAtt on VCI
Learning Nr of 15 min 30 min 60 min Avg. time
Rate Epochs MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
0.020 4.444 7.572 7.177 4.848 8.696 7.949 5.426 10.512 10.482 66.538
0.015 4.444 7.802 7.292 4.800 8.813 8.006 5.437 10.461 9.176 66.919
0.010 100 4.610 8.346 7.750 4.947 9.141 8.269 5.552 10.586 9.201 64.365
0.005 5.075 9.597 8.635 5.435 10.416 9.146 5.965 11.738 9.986 66.625
0.0025 5.211 10.374 9.103 5.536 11.078 9.585 6.013 12.124 10.303 67.116

• Optimizer: Adam (for the Cheb-GRUAtt model) / Adagrad (for the GAT-CNNAtt model)

• Learning rate: 0.0005 (for the Cheb-GRUAtt model) / 0.015 (for the GAT-CNNAtt model)

• Loss function: Mean Squared Error

• Historical Features: 12

• Prediction Horizons: 12

• Number of Spatio-temporal blocks: 2

To make the tables more compact, abbreviations were used to represent the different combina-
tions of weather features, presented in Table 6.17.
The goal was to see if adding additional time features would benefit the results. In theory, this
could help the model to find patterns in the temporal aspect of data, i.e. between the same days
of the week, distinguish between weekdays and weekends and also between different times of the
day. it is important to note that the problem addressed is a short-term prediction, in which some
of these patterns might not have as much influence, so the results should help to answer these
questions.
Empirical Evaluation 64

Table 6.17: Abbreviations for time features to use in the experiments’ tables

Time Features Used Abbreviations

No time features None
Time in day Day
Time in Day + Day in Week Day + Week
Time in Day + Is weekend Day + Weekend
Time in Day + Day in Week + Hour in Day Day + Week + Hour
All time features All

Tables 6.18, 6.19 and 6.20 show the results of adding different combinations of time features
to the input data of the Cheb-GRUAtt model in PeMS-BAY, METR-LA and VCI datasets, respec-
tively.

Table 6.18: Results of Cheb-GRUAtt with the use of different combinations of time features in
PeMS-BAY
15 min 30 min 60 min Avg. time
Time Features
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
None 1.708 3.696 3.407 2.419 5.640 4.866 3.456 8.687 6.615 131.337
Day 1.678 3.673 3.241 2.266 5.231 4.425 2.979 7.214 5.647 130.717
Day + Week 1.688 3.705 3.230 2.233 5.207 4.375 2.909 7.109 5.533 132.326
Day + Weekend 1.692 3.651 3.234 2.255 5.190 4.389 2.931 7.086 5.568 132.010
Day + Week + Hour 1.696 3.724 3.240 2.255 5.252 4.400 2.936 7.159 5.567 133.132
All 1.692 3.679 3.233 2.667 5.230 4.388 2.946 7.119 5.558 130.956

Table 6.19: Results of Cheb-GRUAtt with the use of different combinations of time features in
METR-LA
15 min 30 min 60 min Avg. time
Time Features
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
None 2.850 7.720 5.565 3.554 10.477 6.924 4.695 14.867 8.666 81.829
Day 2.877 7.832 5.500 3.535 10.349 6.748 4.523 14.159 8.266 83.439
Day + Week 2.960 7.980 5.466 3.651 10.519 6.692 4.657 14.239 8.173 83.218
Day + Weekend 2.928 7.944 5.477 3.600 10.400 6.692 4.592 14.029 8.149 83.376
Day + Week + Hour 2.977 7.953 5.475 3.678 10.478 6.672 4.698 14.205 8.163 83.556
All 2.976 8.084 5.477 3.646 10.519 6.679 4.673 14.174 8.151 83.015

Table 6.20: Results of Cheb-GRUAtt with the use of different combinations of time features in
VCI
15 min 30 min 60 min Avg. time
Time Features
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
None 3.785 6.432 6.594 4.451 8.378 7.928 5.382 10.963 9.595 123.141
Day 3.800 6.452 6.458 4.397 8.180 7.658 5.210 10.399 9.151 120.941
Day + Week 3.857 6.540 6.525 4.468 8.307 7.771 5.268 10.510 9.302 119.704
Day + Weekend 3.931 6.584 6.598 4.520 8.266 7.777 5.316 10.409 9.230 119.653
Day + Week + Hour 3.836 6.453 6.487 4.464 8.204 7.724 5.324 10.438 9.259 120.032
All 3.963 6.600 6.568 4.571 8.238 7.780 5.397 10.497 9.329 120.362

In PeMS-BAY, it is possible to see that having no time features produces the worst results, but
in terms of the best results, it is difficult to see which combination is best. ’Time in Day + Day
in Week’ is better in most metrics, but ’Time in Day’ and ’Time in Day + Is Weekend’ perform
6.1 Base Graph Neural Network Architecture 65

better in others. In METR-LA, having no time features performs better for the smallest prediction
horizon, but then it starts getting worse than other alternatives. ’Time in Day’ is better in some
cases, but in others, it is ’Time in Day + Is Weekend’. In the case of VCI, having no time features
works better in the smallest prediction horizon, but in the two other horizons, ’Time in Day’ is
better.
It is worth noting that the differences between the results are very small, with differences
sometimes only on the second and third decimal cases, giving them little significance.
The choice here was to continue using ’Time in Day’ as the only time feature with this model.
Tables 6.21, 6.22 and 6.23 show the results of adding different combinations of time features
to the input data of GAT-CNNAtt in PeMS-BAY, METR-LA and VCI datasets, respectively.

Table 6.21: Results of GAT-CNNAtt with the use of different combinations of time features in
PeMS-BAY
15 min 30 min 60 min Avg. time
Time Features
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
None 2.769 6.303 5.281 2.901 6.581 5.497 3.281 7.463 6.035 120.005
Day 2.751 6.567 5.211 2.849 6.766 5.359 3.117 7.393 5.754 121.505
Day + Week 2.687 6.463 5.185 2.805 6.813 5.390 3.089 7.645 5.893 123.764
Day + Weekend 2.707 6.377 5.119 2.812 6.642 5.277 3.081 7.326 5.694 123.445
Day + Week + Hour 2.746 6.561 5.208 2.878 6.881 5.426 3.186 7.609 5.883 124.842
All 2.755 6.359 5.226 2.852 6.864 5.472 3.108 7.615 5.818 126.565

Table 6.22: Results of GAT-CNNAtt with the use of different combinations of time features in
METR-LA
15 min 30 min 60 min Avg. time
Time Features
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
None 4.979 16.239 9.360 5.387 18.062 9.955 6.021 20.916 10.824 53.081
Day 3.683 10.872 6.687 3.968 12.068 7.301 4.530 14.435 8.344 55.184
Day + Week 5.357 15.962 8.733 5.618 16.997 9.135 6.081 18.905 9.808 55.706
Day + Weekend 4.541 13.532 7.989 4.738 14.250 8.354 5.267 15.711 8.985 55.367
Day + Week + Hour 4.594 14.054 7.870 4.982 15.323 8.416 5.644 17.504 8.416 56.513
All 4.324 12.547 7.497 4.603 13.775 8.072 5.137 16.038 9.035 57.432

Table 6.23: Results of GAT-CNNAtt with the use of different combinations of time features in
VCI
15 min 30 min 60 min Avg. time
Time Features
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
None 4.220 7.670 7.174 4.625 8.873 8.057 5.351 10.776 9.421 70-147
Day 4.444 7.802 7.292 4.800 8.813 8.006 5.437 10.461 9.176 66.919
Day + Week 4.855 8.938 8.068 5.261 9.085 8.867 5.998 11.137 10.088 64.421
Day + Weekend 4.463 8.061 7.553 4.844 8.988 8.212 5.526 10.606 9.458 67.075
Day + Week + Hour 5.650 10.523 9.301 5.896 11.072 9.776 6.362 11.974 10.478 63.894
All 5.161 9.197 9.055 5.389 9.867 9.509 5.744 10.936 10.390 65.300

The influence the time features have when used with the three datasets is different. In the
case of PeMS-BAY the results are all very similar, but for the bigger horizons, the ’Time in Day
+ Day in Week’ combination produces better results. In the case of METR-LA, the results show
Empirical Evaluation 66

more difference between the different combinations, and using only ’Time in Day’ produces the
best results. For VCI, the difference between the different combinations is less noticeable than
in METR-LA but more than in PeMS-BAY. Using no time features works best for the smallest
prediction horizon, but for the bigger ones, ’Time in Day’ produces better results in terms of
MAPE and RMSE.
In terms of training time, the addition or removal of time features does not produce a significant
impact.

[Link] Effect of the number of Historical Features and Prediction Horizons

For this experiment, the settings, apart from the learning rate, were the following:

• Number of training epochs: 100

• Optimizer: Adam (for the Cheb-GRUAtt model) / Adagrad (for the GAT-CNNAtt model)

• Learning rate: 0.0005 (for the Cheb-GRUAtt model) / 0.015 (for the GAT-CNNAtt model)

• Loss function: Mean Squared Error

• Number of Spatio-temporal blocks: 2

• Time features: ’Time in Day’

The goal of this experiment was to see the impact that changing the number of historical fea-
tures given as input would change the results for each prediction horizon. Moreover, it was also
to see how the models perform on longer prediction, i.e predicting 90 and 120 minutes ahead. As
explained in Section 5.3.4, having, for example, ’Historic Features’ be 6 and ’Prediction Hori-
zon’ be 3, means that 6 × 5 = 30 minutes before the given timestamp were used as historical
features to predict 3 × 5 = 15 minutes into the future from the given timestamp. Tables 6.25 and
6.24 contain the results of different combinations of historical features and prediction horizons
for Cheb-GRUAtt and GAT-CNNAtt, respectively. The reason this study was only made on one
dataset, comes from the fact that training the models especially when there are more historic fea-
tures is very slow and it requires a lot of computational time to be able to extract the results to fill
the tables. METR-LA is the dataset in which the models train faster, so it was chosen as the only
one where this would be tried.
Some conclusions can be extracted from observing the tables:

• By comparing the results on the same prediction horizon, increasing the number of historical
features makes the results worse. Comparing the results of prediction up to 60 minutes (12
prediction steps), the results when using 12, 18 or 24 historical features get progressively
worse results, and this applies to the other prediction horizons.

• By comparing the use of the same number of historical features for different numbers of
prediction horizons, the results, in general, get better the more horizons are predicted.
6.1 Base Graph Neural Network Architecture 67

Table 6.24: Results of Cheb-GRUAtt with different numbers of historical features and output
predictions on METR-LA dataset
Hist. 15 min 30 min 60 min 90 min 120 min Time
Pred.
Feat. MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE (s)
3 3 2.903 7.821 5.491 - - - - - - - - - - - - 21.041
6 3 3.573 10.620 6.796 - - - - - - - - - - - - 21.011
6 6 2.893 7.795 5.480 3.555 10.490 6.746 - - - - - - - - - 41.900
12 3 4.529 14.278 8.257 - - - - - - - - - - - - 21.192
12 6 4.126 12.670 7.592 4.563 14.320 8.227 - - - - - - - - - 42.133
12 12 2.902 7.944 5.496 3.544 10.434 6.734 4.493 14.136 8.237 - - - - - - 83.439
18 3 5.287 17.046 9.192 - - - - - - - - - - - - 21.341
18 6 4.900 15.667 8.731 5.225 16.853 9.157 - - - - - - - - - 42.169
18 12 4.090 12.578 7.593 4.506 14.178 8.209 5.205 16.911 9.164 - - - - - - 83.202
18 18 2.935 7.951 5.583 3.573 10.523 6.860 4.552 14.901 8.581 5.342 17.391 9.304 - - - 123.031
24 3 5.934 19.196 9.917 - - - - - - - - - - - - 21.774
24 6 5.577 18.028 9.552 5.813 18.938 9.855 - - - - - - - - - 42.368
24 12 4.979 15.673 8.769 5.292 16.863 9.173 5.859 18.991 9.847 - - - - - - 83.283
24 18 4.143 12.765 7.634 4.567 14.345 8.255 5.267 17.357 9.332 5.817 19.049 9.894 - - - 124.000
24 24 2.951 8.058 5.561 3.608 10.523 6.797 4.725 14.687 8.474 5.426 17.367 9.395 5.908 19.206 9.976 174.071

Table 6.25: Results of GAT-CNNAtt with different numbers of historical features and output pre-
dictions on METR-LA dataset
Hist. 15 min 30 min 60 min 90 min 120 min Time
Pred.
Feat. MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE (s)
3 3 3.497 10.051 6.683 - - - - - - - - - - - - 26.187
6 3 3.810 11.996 7.163 - - - - - - - - - - - - 35.618
6 6 3.903 11.862 7.055 4.178 12.980 7.672 - - - - - - - - - 34.248
12 3 5.327 18.657 9.43 - - - - - - - - - - - - 56.874
12 6 4.723 14.691 8.652 4.951 15.500 9.022 - - - - - - - - - 55.239
12 12 3.683 10.872 6.687 3.968 12.068 7.301 4.530 14.435 8.344 - - - - - - 55.356
18 3 4.753 14.136 8.762 - - - - - - - - - - - - 77.584
18 6 4.445 13.903 8.006 4.635 14.629 8.424 - - - - - - - - - 75.963
18 12 4.883 15.808 8.858 4.983 16.190 9.067 5.318 17.458 9.612 - - - - - - 76.270
18 18 5.646 18.698 9.672 5.737 19.090 9.872 6.077 20.314 10.393 6.357 21.368 10.749 - - - 76.004
24 3 6.682 24.200 11.8739 - - - - - - - - - - - - 98.448
24 6 5.369 17.501 9.819 5.504 18.026 10.058 - - - - - - - - - 94.667
24 12 5.921 20.211 10.755 6.105 20.904 11.080 6.567 22.471 11.797 - - - - - - 96.765
24 18 5.523 17.527 9.624 5.621 18.102 9.779 5.863 19.232 10.100 6.140 20.369 6.140 - - - 96.172
24 24 6.197 20.945 10.594 6.349 21.673 10.840 6.605 22.949 11.258 6.803 23.976 11.605 6.950 24.777 11.891 96.412

• As seen in the other experiments, when observing a single row (results across different
horizons for the same configuration), the results get progressively worse as the horizons get
bigger.

• The results obtained with Cheb-GRUAtt are better than GAT-CNNAtt, across all combina-
tions in all metrics.

• When looking at the bigger horizons of 90 and 120 minutes, it is possible to see that the
results achieved by the models get very degraded, which shows that their applicability is
focused towards short-term prediction, as also pointed out by the recent literature. It is still
possible to see, however, that the Cheb-GRUAtt model has better results and shows less
degradation than the GAT-CNNAtt model.

6.1.3 Effect of the Length of the Training Set

With the retrieval of the PeMS-BAY-2years dataset, it was possible to test if training on a longer
period of recorded data would produce better results. For this, 4 different training were carried
out: first using only the 6 most recent months, then the 12 more recent, then 18 and finally the
whole 24 months. It’s important to note that the test set for evaluating the 4 different sets was kept
Empirical Evaluation 68

the same. Given training time restrictions, only the Cheb-GRUAtt Model was experimented on.
The settings for this experiment were the following:

• Number of training epochs: 100

• Optimizer: Adam

• Learning rate: 0.0005

• Loss function: Mean Squared Error

• Number of Spatio-temporal blocks: 2

• Time features: ’Time in Day’

• Historical Features: 12

• Prediction Horizons: 12

Table 6.26: Comparison of the results with different training dataset lengths

Dataset 15 min 30 min 60 min Avg. time

Length MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE per Epoch (s)
6 months 1.280 2.562 2.603 1.634 3.208 3.432 2.046 4.517 4.202 122.346
12 months 1.198 2.402 2.561 1.524 3.201 3.346 1.890 4.097 4.080 255.478
18 months 1.280 2.556 2.602 1.658 3.473 3.421 2.032 4.453 4.171 370.385
24 months 1.234 2.406 2.563 1.592 3.253 3.375 2.110 4.401 4.164 508.349

Table 6.26 shows the results of the different dataset lengths used for training. The results show
that increasing the dataset length to 12 months produces better results. This can come from the
fact that training with a full year allows the model to capture the full extent of the seasonal de-
pendencies. When extending the results further to 18 months, go back to being similar to training
with 6 months. Training with 24 months is slightly better than with 18 but worse than with 12
months. These results indicate that training with one year of data can help the results, but further
increasing the training length is not favourable. When looking at the training times, they show a
linear increase in time with the length increase, i.e. training with 12 months takes roughly twice
the time as training with 6 months and so on. This lack of better results with the increase in dataset
length could come from the fact that the less recent data present in those parts happened in a time
that was more affected by COVID, which impacted the typical transportation patterns.

6.2 Handling Missing Data

Now that the experiments with the base GNN approach are tackled, the results when using missing
data imputation will be explored. From now on, since Cheb-GRUAtt got the best results in all
experiments when compared to GAT-CNNAtt, the experiments will only be performed on Cheb-
GRUAtt, given the computational time limitations.
For this experiment, the settings were the following:
6.2 Handling Missing Data 69

• Number of training epochs: 100

• Optimizer: Adam

• Learning rate: 0.0005

• Loss function: Mean Squared Error

• Number of Spatio-temporal blocks: 2

• Time features: ’Time in Day’

• Historical Features: 12

• Prediction Horizons: 12

Table 6.27 shows the results of using different missing data imputation techniques on the
PeMS-BAY dataset.
Table 6.27: Comparison of the results with the different missing data imputation techniques in the
PeMS-BAY dataset
15 min 30 min 60 min Avg. time
Techniques
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE per Epoch (s)
No imputation 1.678 3.673 3.241 2.266 5.231 4.425 2.979 7.214 5.647 130.717
Mean 1.678 3.680 3.242 2.273 5.253 4.426 2.998 7.248 5.647 131.921
LOCF+NOCB 1.678 3.680 3.243 2.265 5.262 4.426 2.997 7.249 5.647 131.257
Interpolation 1.678 3.679 3.242 2.273 5.253 4.426 2.997 7.247 5.647 131.678
MICE 1.688 3.705 3.260 2.288 5.289 4.459 3.042 7.404 5.728 132.002
MICE+interpolation 1.678 3.679 3.242 2.273 5.253 4.426 2.997 7.247 5.647 131.976

The results here show very little difference between the different techniques and even worsen
them in the case of MICE. Since this particular dataset has very little missing data, the impact they
have on the results is minimal, so using the missing data imputation techniques did not make a
noticeable different. This is why no results were highlighted in the table.
Table 6.28 shows the results of using different missing data imputation techniques on the
METR-LA dataset.
Table 6.28: Comparison of the results with the different missing data imputation techniques in the
METR-LA dataset
15 min 30 min 60 min Avg. time
Techniques
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE per Epoch (s)
No imputation 2.877 7.832 5.500 3.535 10.349 6.748 4.523 14.159 8.266 83.439
Mean 2.925 7.486 5.503 3.547 9.538 6.318 4.446 12.481 8.006 83.196
LOCF+NOCB 2.840 7.799 5.486 3.501 10.312 6.700 4.491 14.122 8.221 83.345
Interpolation 2.862 7.665 5.417 3.504 10.004 6.626 4.464 13.500 8.115 83.542
MICE 2.861 7.598 5.531 3.583 9.674 6.679 4.490 12.672 8.054 83.630
MICE+interpolation 2.862 7.665 5.417 3.504 10.004 6.626 4.464 13.500 8.114 83.965

In this dataset, the results show a bit more difference. Using the mean, despite its simplicity
and not being the preferred technique for time series data imputation, showed the best results,
although closely followed most of the time by MICE.
Empirical Evaluation 70

Table 6.29 shows the results when used with VCI.

Table 6.29: Comparison of the results with the different missing data imputation techniques in the
VCI dataset
15 min 30 min 60 min Avg. time
Techniques
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE per Epoch (s)
No imputation 3.800 6.452 6.458 4.397 8.180 7.658 5.210 10.399 9.151 120.941
Mean 3.778 6.338 6.437 4.386 8.129 7.665 5.225 10.451 9.168 123.976
LOCF+NOCB 3.887 6.395 6.522 4.486 8.200 7.704 5.282 10.393 9.181 122.234
Interpolation 3.857 6.373 6.508 4.452 8.102 7.693 5.269 10.385 9.157 120.869
MICE 3.825 6.890 6.430 4.417 8.174 7.656 5.221 10.367 9.119 121.210
MICE+interpolation 3.856 6.442 6.493 4.450 8.179 7.707 5.260 10.402 9.191 124.639

Here the results of the different techniques show a bit more influence than in PeMS-BAY but
less than in METR-LA. This can stem from the fact that PeMS-BAY has the least amount of miss-
ing data, followed by VCI and then METR-LA. Having less missing data makes the imputation
used to fill them less impactful on the overall results. Given the results, it can be understood
why most of the previous works on GNNs for traffic forecasting don’t pay a lot of attention to
the missing data, given that the models, especially when their occurrence is small, can deal well
with their presence. But since the results show that more impact is noticed when the presence of
missing data is bigger, meaning that the models are not immune to missing data and can benefit
from the use of these techniques. Furthermore, this shows that when dealing with datasets where
there are even more missing data than in METR-LA, the usage of an imputation technique can
have an even more significant impact and should not be overlooked. One detail worth mentioning
is that, when comparing the imputation techniques that performed better pre-process the data for
the GNN models with the results shown in Table 5.7, where the power of imputation is evaluated
on its own, the results don’t match up. That is, the techniques that were better at generating values
closer to what they would really be, where not the ones that helped the model achieve the best
results. This is seen especially in the mean imputation, which did not perform well on its own
as imputation but when used as pre-processing for the GNN models, helped them achieve better
results.

6.3 Weather Data Incorporation

For this experiment, the settings were the following:

• Number of training epochs: 100

• Optimizer: Adam

• Learning rate: 0.0005

• Loss function: Mean Squared Error

• Number of Spatio-temporal blocks: 2

6.3 Weather Data Incorporation 71

• Time features: ’Time in Day’

• Missing Data imputation technique: None

To make the tables more compact, abbreviations were used to represent the different combina-
tions of weather features, presented in Table 6.30

Table 6.30: Abbreviations for weather features to use in the experiments’ tables

Weather Features Used Abbreviations

No weather features None
precip_rate + visibility p+v
precip_rate + visibility + wind_spd p+v+w
precip_rate + wind_spd p+w
precip_rate + wind_spd + visibility + temp p+w+v+t
All features All

Tables 6.31, 6.32 and 6.33 show the results obtained when adding the weather features to the
input of the Cheb-GRUAtt model in PeMS-BAY, METR-LA and VCI, respectively.

Table 6.31: Results with the use of different combinations of weather features in PeMS-BAY
15 min 30 min 60 min Avg. time
Weather Features
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
None 1.678 3.673 3.241 2.266 5.231 4.425 2.979 7.214 5.647 130.717
p+v 1.691 3.746 3.256 2.277 5.320 4.438 2.986 7.314 5.676 132.992
p+v+w 1.781 3.984 3.530 2.412 5.767 4.885 3.396 8.647 6.548 132.659
p+w 1.696 3.726 3.247 2.285 5.294 4.431 2.989 7.270 5.673 132.973
p+w+v+t 1.736 3.899 3.304 2.343 5.482 4.499 3.093 7.501 5.761 132.638
All 1.749 3.824 3.280 2.349 5.389 4.472 3.090 7.393 5.744 134.078

Table 6.32: Results with the use of different combinations of weather features in METR-LA
15 min 30 min 60 min Avg. time
Weather Features
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
None 2.877 7.832 5.500 3.535 10.349 6.748 4.523 14.159 8.266 83.439
p+v 2.893 7.882 5.517 3.570 10.451 6.781 4.593 14.306 8.310 83.222
p+v+w 2.907 7.932 5.507 3.569 10.444 6.756 4.590 14.307 8.280 85.292
p+w 2.891 7.886 5.493 3.572 10.433 6.752 4.583 14.306 8.295 83.863
p+w+v+t 2.899 8.035 5.512 3.575 10.587 6.772 4.567 14.353 8.306 83.414
All 2.986 8.102 5.511 3.667 10.649 6.754 4.661 14.370 8.338 83.257

Table 6.33: Results with the use of different combinations of weather features in VCI
15 min 30 min 60 min Avg. time
Weather Features
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE p/ Epoch (s)
None 3.800 6.452 6.458 4.397 8.180 7.658 5.210 10.399 9.151 120.941
p+v 3.845 6.539 6.495 4.458 8.249 7.722 5.273 10.348 9.179 123.201
p+v+w 3.860 6.540 6.532 4.470 8.286 7.756 5.293 10.544 9.252 118.959
p+w 3.866 6.585 6.540 4.489 8.362 7.815 5.365 10.614 9.371 123.039
p+w+v+t 3.926 6.741 6.603 4.558 8.554 7.873 5.431 10.841 9.436 119.066
All 3.992 6.802 6.635 4.658 8.745 7.906 5.431 10.872 9.546 120.363
Empirical Evaluation 72

The results observed in the three datasets are similar in the sense that adding the features does
not improve the results and instead makes them slightly worse. The combinations that almost
have the same performance as having no features are ’precip_rate + visibility’ and ’precip_rate
+ wind_spd’. The lack of good results could mean that since the problem is a short-term traffic
prediction, these weather dependencies don’t have as much influence as in longer-term predictions.
In terms of training time required, adding these features does not add a substantial amount of
time. This shows that despite the increase in the input data size, the models take roughly the same
time to process it.

6.4 Global Comparisons

After making the comparisons between different approaches within the scope of the work, it makes
sense to finally present how the results compare against other techniques. Tables 6.34, 6.35 and
6.36 show the results of comparing the developed approaches with baseline techniques on the
three datasets: PeMS-BAY, METR-LA and VCI, respectively. In VCI the GNN-based techniques
STGCN and DCRNN are not displayed since the results provided are the ones stated by the papers
themselves, and because of that, there are no results for VCI. In an ideal scenario, the results with
these two techniques should have been trained and tested in this work, but the code available for
them was running on outdated software versions, and it was not possible to get them to work as
expected originally.

Table 6.34: Comparison of the best-performing models with baselines in PeMS-BAY

15 min 30 min 60 min

Techniques
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE
HA 3.401 8.169 6.693 3.401 8.169 6.693 3.401 8.169 6.693
ARIMA 1.702 3.564 3.325 2.367 5.506 4.798 3.389 8.452 6.549
VAR 1.786 3.874 3.179 2.356 5.125 4.323 2.935 6.764 5.617
DCRNN 1.380 2.900 2.950 1.740 3.900 3.970 2.070 4.900 4.740
STGCN 1.360 2.900 2.960 1.810 4.170 4.270 2.490 5.790 5.690
ChebGRUAtt 1.688 3.705 3.230 2.233 5.207 4.375 2.909 7.109 5.533

Table 6.35: Comparison of the best-performing models with baselines in METR-LA

15 min 30 min 60 min
Techniques
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE
HA 11.013 23.340 14.737 11.013 23.340 14.737 11.013 23.340 14.737
ARIMA 3.990 9.600 8.210 5.150 12.700 10.450 6.900 17.400 13.230
VAR 4.420 10.200 7.890 5.410 12.700 10.450 6.520 10.110 15.800
DCRNN 2.770 7.300 5.380 3.150 8.800 6.450 3.600 10.500 7.590
STGCN 2.870 7.400 5.540 3.480 9.400 6.840 4.450 11.800 8.410
ChebGRUAtt 2.925 7.486 5.503 3.547 9.538 6.318 4.446 12.481 8.006
ChebGRUAtt + Missing Data 2.877 7.832 5.500 3.535 10.349 6.748 4.523 14.159 8.266

In PeMS-BAY the results of the developed models fall short of the expectations compared to
the other approaches. ChebGRUAtt has worse results than the other GNN-based approaches in
6.5 Results Discussion 73

Table 6.36: Comparison of the best-performing models with baselines in VCI

15 min 30 min 60 min

Techniques
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE
HA 9.234 17.652 12.109 9.234 17.652 12.109 9.234 17.652 12.109
ARIMA 4.283 9.781 8.346 5.097 12.540 10.002 6.764 15.458 13.166
VAR 4.443 10.010 8.590 5.410 12.626 10.237 6.638 15.295 13.450
ChebGRUAtt 3.785 6.432 6.594 4.451 8.378 7.928 5.382 10.963 9.595
ChebGRUAtt + Missing Data 3.825 6.890 6.430 4.417 8.174 7.656 5.221 10.367 9.119

almost all cases and even when compared to VAR in some instances. Since the results of PeMS-
BAY are much better overall than in the other datasets, the room for improvement is smaller, and
time-series techniques have results more on par with the GNN-based techniques. This can be from
the lack of missing data on this dataset or the overall quality of the data present in it. In datasets
like this, where time-series data performs well, it actually can trigger the question of whether
deep-learning-based techniques (such as GNNs) are worth employing, given that the computation
power and time required for training these models is substantial, especially when compared with
the time-series algorithms employed here.
In METR-LA however, it outperforms all time-series-based approaches in all metrics and hori-
zons, and it also outperforms STGCN in some cases. DCRNN, however still has the best results
overall. In terms of VCI, it is only possible to compare with the time-series-based approaches, and
it is significantly better than those for all metrics and horizons.

6.5 Results Discussion

In sum, the empirical experiments explored and answered several different questions:

• What combinations of spatial and temporal mechanisms work best for this problem?
It was found that the combination of Chebyshev Spectral Graph Convolution and Gated
Recurrent Unit was the best model across all three datasets. It was possible to see the way
the explored models perform relative to each other was the same across all datasets, i.e. if
model X is better than model Y in dataset Z, it also is better in dataset W.

• What hyperparameter settings are better?

It was found that they work differently in different models, but within the same model, the
results when using the same hyperparameter settings in other datasets worked more or less
in the same way.

• Do missing data imputation techniques help achieve better results?

what was discovered is that the impact the measure have depends on the amount of missing
data occurring in the original dataset. The effect can be close to null if the amount is min-
imal. However, when the presence is more significant, the results improve with the use of
imputation techniques, especially when using mean imputation and MICE.
Empirical Evaluation 74

• Can the models extract helpful knowledge from adding weather features and get better re-
sults?
Their impact on the results was negative, pointing against the use of those features.

• How do the models compare to existing techniques?

Compared to other techniques, the results achieved are not groundbreaking, but they point
in the direction that GNNs are a good approach to this traffic prediction problem, and they
can be used in further research of other mechanisms and hyperparameter configurations.

Something worth mentioning that could be seen throughout the chapter is that time constraints
impacted the number of experiments carried out and the different combinations that could have
been explored. There is the possibility that some of the combinations that were not explored could
have led to better results than the ones presented, even if the tests chosen to do were the most
theoretically promising. This is a common struggle in studying machine learning algorithms, and
despite trying to minimize its impact by using multiple machines to run the experiences, it still
posed an obstacle.
Chapter 7

Conclusions and Future Work

The motivation behind this research stems from the increasing demand for reliable and accurate
traffic predictions in today’s fast-paced world. With the increasing number of vehicles on the
road, traffic congestion is becoming a widespread issue, leading to higher levels of pollution and
decreased mobility. To address this problem, the field of traffic forecasting has been subject to
thorough research over the years, given the complexity of the spatial and temporal dependencies
of traffic data.
Approaches to traffic forecasting include various techniques that range from traditional statis-
tical methods to Deep Learning. However, most approaches have their limitations when it comes
to handling the complex temporal and spatial dependencies of graph-like traffic data. This is where
Graph Neural Networks come into the picture. GNNs have been shown to be effective in handling
complex relationships between data and have been successfully applied to various problems.
The objective of this dissertation consisted of applying GNNs to the problem of traffic forecast-
ing and evaluating their performance against other frequently used methods, considering missing
data handling and the incorporation of external factors as additional methods to try to improve
results.
An important contribution of this work was the standardization of the process of generating
the input data for the GNN models, in order to easily allow for the training and testing in different
datasets.
The empirical evaluation showed that the explored GNNs outperform several commonly used
time-series-based models in the literature, affirming the leverage that these models can have. How-
ever, they did not outperform other GNN-based techniques existing in the literature, meaning that
some more tuning to the architectures and parameters should be done.
Another contribution of this dissertation is the study of the impact of using missing data im-
putation techniques on prediction performance. It was shown that GNNs can benefit from these
techniques, especially when used with VCI and METR-LA datasets. The research also demon-
strated that the incorporation of weather conditions does not produce a positive impact on the
results, contrary to the initial hypothesis.

75
Conclusions and Future Work 76

Moving forward, the potential applications of GNNs in Intelligent Transportation Systems are
vast, and further research is needed to explore their full potential. This work opens several direc-
tions for future work, such as a more intensive study of hyperparameters and layer architectures to
try to achieve better results.
On top of that, a broader study of sources of external data, such as social events happening
close to the road network, social media mentions of congestion or even accidents happening on
the network could be used to enrich the models.
References

Alam, I., Ahmed, M. F., Alam, M., Ulisses, J., Farid, D. M., Shatabda, S., & Rossetti, R. J. F.
(2017). Pattern mining from historical traffic big data. In 2017 ieee region 10 symposium
(tensymp) (p. 1-5). doi: 10.1109/TENCONSpring.2017.8070031
Alam, I., Farid, D. M., & Rossetti, R. J. F. (2019). The prediction of traffic flow with regres-
sion analysis. In A. Abraham, P. Dutta, J. K. Mandal, A. Bhattacharya, & S. Dutta (Eds.),
Emerging technologies in data mining and information security (pp. 661–671). Singapore:
Springer Singapore.
Alsolami, B., Mehmood, R., & Albeshri, A. (2020). Hybrid Statistical and Machine Learning
Methods for Road Traffic Prediction: A Review and Tutorial. In R. Mehmood, S. See,
I. Katib, & I. Chlamtac (Eds.), Smart Infrastructure and Applications: Foundations for
Smarter Cities and Societies (pp. 115–133). Springer International Publishing. doi: 10
.1007/978-3-030-13705-2_5
Ang, K. L.-M., Seng, J. K. P., Ngharamike, E., & Ijemaru, G. K. (2022). Emerging Technologies
for Smart Cities&rsquo; Transportation: Geo-Information, Data Analytics and Machine
Learning Approaches. ISPRS International Journal of Geo-Information, 11(2). doi: 10
.3390/ijgi11020085
Angarita-Zapata, J. S., Masegosa, A. D., & Triguero, I. (2019). A taxonomy of traffic forecasting
regression problems from a supervised learning perspective. IEEE Access, 7, 68185-68205.
doi: 10.1109/ACCESS.2019.2917228
Baldassarre, F., & Azizpour, H. (2019). Explainability techniques for graph convolutional net-
works. ArXiv, abs/1905.13686.
Barros, J., Araujo, M., & Rossetti, R. J. F. (2015, June). Short-term real-time traffic pre-
diction methods: A survey. In 2015 International Conference on Models and Tech-
nologies for Intelligent Transportation Systems (MT-ITS) (pp. 132–139). doi: 10.1109/
MTITS.2015.7223248
Bhanja, S., & Das, A. (2019, January). Impact of Data Normalization on Deep Neural Network
for Time Series Forecasting. arXiv. doi: 10.48550/arXiv.1812.05519
Bokaba, T., Doorsamy, W., & Paul, B. S. (2022, January). A Comparative Study of Ensemble
Models for Predicting Road Traffic Congestion. Applied Sciences, 12(3), 1337. doi: 10
.3390/app12031337
Bui, K.-H. N., Cho, J., & Yi, H. (2022, February). Spatial-temporal graph neural network for traffic

77
References 78

forecasting: An overview and open research issues. Applied Intelligence, 52(3), 2763–2774.
doi: 10.1007/s10489-021-02587-w
Cai, L., Zhang, Z., Yang, J., Yu, Y., Zhou, T., & Qin, J. (2019, December). A noise-immune
Kalman filter for short-term traffic flow forecasting. Physica A: Statistical Mechanics and
its Applications, 536, 122601. doi: 10.1016/[Link].2019.122601
Cao, S., Wu, L., Zhang, R., Li, J., & Wu, D. (2022, July). Capturing Local and Global
Spatial-Temporal Correlations of Spatial-Temporal Graph Data for Traffic Flow Predic-
tion. In 2022 International Joint Conference on Neural Networks (IJCNN) (pp. 1–8). doi:
10.1109/IJCNN55064.2022.9892616
Cao, W., Wang, D., Li, J., Zhou, H., Li, L., & Li, Y. (2018). BRITS: Bidirec-
tional Recurrent Imputation for Time Series. In Advances in Neural Informa-
tion Processing Systems (Vol. 31). Curran Associates, Inc. Retrieved 2023-
01-31, from [Link]
[Link]
Castro, P. S., Zhang, D., & Li, S. (2012). Urban Traffic Modelling and Prediction Using Large
Scale Taxi GPS Traces. In J. Kay, P. Lukowicz, H. Tokuda, P. Olivier, & A. Krüger (Eds.),
Pervasive Computing (pp. 57–72). Berlin, Heidelberg: Springer. doi: 10.1007/978-3-642
-31205-2_4
Che, Z., Purushotham, S., Cho, K., Sontag, D. A., & Liu, Y. (2016). Recurrent neural networks
for multivariate time series with missing values. CoRR, abs/1606.01865. Retrieved from
[Link]
Chen, C., Li, K., Teo, S. G., Zou, X., Wang, K., Wang, J., & Zeng, Z. (2019, July). Gated
Residual Recurrent Graph Neural Networks for Traffic Prediction. Proceedings of the AAAI
Conference on Artificial Intelligence, 33(01), 485–492. doi: 10.1609/aaai.v33i01.3301485
Chen, D., Yan, X., Liu, X., Li, S., Wang, L., & Tian, X. (2021). A multiscale-grid-based stacked
bidirectional gru neural network model for predicting traffic speeds of urban expressways.
IEEE Access, 9, 1321-1337. doi: 10.1109/ACCESS.2020.3034551
Chen, X., Cai, X., Liang, J., & Liu, Q. (2018). Ensemble learning multiple lssvr with improved
harmony search algorithm for short-term traffic flow forecasting. IEEE Access, 6, 9347-
9357. doi: 10.1109/ACCESS.2018.2805299
Chen, X., Chen, Y., & He, Z. (2018, March). Urban traffic speed dataset of guangzhou, china.
Zenodo. doi: 10.5281/zenodo.1205229
Cui, Z. (2023, January). Seattle Inductive Loop Detector Dataset V.1 (2015). Retrieved 2023-01-
30, from [Link]
Cui, Z., Henrickson, K. C., Ke, R., & Wang, Y. (2018). Traffic graph convolutional recurrent neural
network: A deep learning framework for network-scale traffic learning and forecasting.
IEEE Transactions on Intelligent Transportation Systems, 21, 4883-4894. doi: 10.48550/
arXiv.1802.07007
Defferrard, M., Bresson, X., & Vandergheynst, P. (2016). Convolutional neural networks on
graphs with fast localized spectral filtering. CoRR. doi: 10.48550/arXiv.1606.09375
References 79

Ge, L., Li, H., Liu, J., & Zhou, A. (2019, June). Temporal Graph Convolutional Networks for
Traffic Speed Prediction Considering External Factors. In 2019 20th IEEE International
Conference on Mobile Data Management (MDM) (pp. 234–242). doi: 10.1109/MDM.2019
.00-52
Gehring, J., Auli, M., Grangier, D., Yarats, D., & Dauphin, Y. N. (2017, July). Convolu-
tional Sequence to Sequence Learning. CoRR, 1243–1252. Retrieved 2023-06-19, from
[Link]
Guo, S., Lin, Y., Feng, N., Song, C., & Wan, H. (2019, July). Attention Based Spatial-Temporal
Graph Convolutional Networks for Traffic Flow Forecasting. Proceedings of the AAAI Con-
ference on Artificial Intelligence, 33(01), 922–929. Retrieved from [Link]
.org/[Link]/AAAI/article/view/3881 doi: 10.1609/aaai.v33i01.3301922
Hamilton, W. L. (2020). Graph Representation Learning. Cham: Springer International Publish-
ing. doi: 10.1007/978-3-031-01588-5
He, Z., Chow, C.-Y., & Zhang, J.-D. (2019, June). STCNN: A Spatio-Temporal Convolutional
Neural Network for Long-Term Traffic Prediction. In 2019 20th IEEE International Confer-
ence on Mobile Data Management (MDM) (pp. 226–233). doi: 10.1109/MDM.2019.00-53
Hu, J., Lin, X., & Wang, C. (2022, July). DSTGCN: Dynamic Spatial-Temporal Graph Convo-
lutional Network for Traffic Prediction. IEEE Sensors Journal, 22(13), 13116–13124. doi:
10.1109/JSEN.2022.3176016
Irawan, K., Yusuf, R., & Prihatmanto, A. S. (2020, December). A Survey on Traffic Flow Predic-
tion Methods. In 2020 6th International Conference on Interactive Digital Media (ICIDM)
(pp. 1–4). doi: 10.1109/ICIDM51048.2020.9339675
Ismiguzel, I. (2022, May). Imputing Missing Data with Simple and Advanced Techniques.
Retrieved 2023-06-10, from [Link]
-missing-data-with-simple-and-advanced-techniques-f5c7b157fb87
Jensen, M. (2019, August). Data Leakage and how to avoid it. Retrieved 2023-
06-12, from [Link]
-to-avoid-it/
Jiang, R., Yin, D., Wang, Z., Wang, Y., Deng, J., Liu, H., . . . Shibasaki, R. (2021). Dl-
traff: Survey and benchmark of deep learning models for urban traffic prediction. In Pro-
ceedings of the 30th acm international conference on information & knowledge manage-
ment (p. 4515–4525). New York, NY, USA: Association for Computing Machinery. doi:
10.1145/3459637.3482000
Jiang, W., & Luo, J. (2021). Big data for traffic estimation and prediction: A survey of data and
tools. CoRR, abs/2103.11824. doi: 10.48550/arXiv.2103.11824
Jiang, W., & Luo, J. (2022, November). Graph neural network for traffic forecasting: A survey.
Expert Systems with Applications, 207, 117921. doi: 10.1016/[Link].2022.117921
Karunasingha, D. S. K. (2022, March). Root mean square error or mean absolute error? Use their
ratio as well. Information Sciences, 585, 609–629. doi: 10.1016/[Link].2021.11.036
References 80

Kipf, T. N., & Welling, M. (2017, February). Semi-Supervised Classification with Graph Convo-
lutional Networks. arXiv. doi: 10.48550/arXiv.1609.02907
Kołodziej, J., Hopmann, C., Coppa, G., Grzonka, D., & Widłak, A. (2022). Intelligent Trans-
portation Systems – Models, Challenges, Security Aspects. In J. Kołodziej, M. Repetto,
& A. Duzha (Eds.), Cybersecurity of Digital Service Chains: Challenges, Methodolo-
gies, and Tools (pp. 56–82). Cham: Springer International Publishing. doi: 10.1007/
978-3-031-04036-8_3
Lau, S. (2017, August). Learning Rate Schedules and Adaptive Learning Rate Methods for
Deep Learning. Retrieved 2023-03-28, from [Link]
learning-rate-schedules-and-adaptive-learning-rate-methods-for
-deep-learning-2c8f433990d1
Laña, I., Del Ser, J., Velez, M., & Vlahogianni, E. (2018, July). Road Traffic Forecasting: Recent
Advances and New Challenges. IEEE Intelligent Transportation Systems Magazine, 10,
93–109. doi: 10.1109/MITS.2018.2806634
Lee, K., Eo, M., Jung, E., Yoon, Y., & Rhee, W. (2021). Short-term traffic prediction with deep
neural networks: A survey. IEEE Access, 9, 54739-54756. doi: 10.1109/ACCESS.2021
.3071174
Lee, K., & Rhee, W. (2022). DDP-GCN: Multi-graph convolutional network for spatiotemporal
traffic forecasting. Transportation Research Part C: Emerging Technologies, 134, 103466.
doi: [Link]
Lee, S., & b. Fambro, D. (1999). Application of subset autoregressive integrated moving average
model for short-term freeway traffic volume forecasting. Transportation Research Record,
1678, 179 - 188.
Li, Y., Mensi, F., & Yu, R. (2023, June). Diffusion Convolutional Recurrent Neural Network:
Data-Driven Traffic Forecasting. Retrieved 2023-06-06, from [Link]
liyaguang/DCRNN
Li, Y., Yu, R., Shahabi, C., & Liu, Y. (2018). Diffusion convolutional recurrent neural network:
Data-driven traffic forecasting. In International conference on learning representations. doi:
10.48550/arXiv.1707.01926
Liu, J., Wu, N., Qiao, Y., & Li, Z. (2021, 01). A scientometric review of research on traffic fore-
casting in transportation. IET Intelligent Transport Systems, 15. doi: 10.1049/itr2.12024
Liu, Z., & Zhou, J. (2020). Introduction to Graph Neural Networks. Cham: Springer International
Publishing. doi: 10.1007/978-3-031-01587-8
Lopes, J., Bento, J., Huang, E., Antoniou, C., & Ben-Akiva, M. (2010, September). Traffic and
mobility data collection for real-time applications. In 13th International IEEE Conference
on Intelligent Transportation Systems (pp. 216–223). doi: 10.1109/ITSC.2010.5625282
Lu, H., Huang, D., Song, Y., Jiang, D., Zhou, T., & Qin, J. (2020, Septembet). St-trafficnet: A
spatial-temporal deep learning network for traffic forecasting. Electronics. doi: 10.3390/
electronics9091474
Lu, J., Li, B., Li, H., & Al-Barakani, A. (2021). Expansion of city scale, traffic modes, traffic
References 81

congestion, and air pollution. Cities, 108, 102974. doi: 10.1016/[Link].2020.102974

Ma, T., Antoniou, C., & Toledo, T. (2020). Hybrid machine learning algorithm and statistical time
series model for network-wide traffic forecast. Transportation Research Part C-emerging
Technologies, 111, 352-372.
Ma, X., Dai, Z., He, Z., Ma, J., Wang, Y., & Wang, Y. (2017). Learning traffic as images: A
deep convolutional neural network for large-scale transportation network speed prediction.
Sensors (Basel, Switzerland), 17.
Mori, U., Mendiburu, A., Álvarez, M., & Lozano, J. A. (2015). A review of travel time estimation
and forecasting for Advanced Traveller Information Systems. Transportmetrica A Transport
Science, 11(2), 119–157. doi: [Link]
O’Connor, R. (2021, December). PyTorch vs TensorFlow in 2023. Retrieved 2023-03-27, from
[Link]
Pavlyuk, D. (2017, January). Short-term Traffic Forecasting Using Multivariate Autoregressive
Models. Procedia Engineering, 178, 57–66. doi: 10.1016/[Link].2017.01.062
PredictHQ. (2023a). Events api. [Link] Retrieved 2023-02-16, from
[Link]
PredictHQ. (2023b). Events - Technical Documentation - PredictHQ. Retrieved 2023-02-16, from
[Link]
Raschka, S., Patterson, J., & Nolet, C. (2020, April). Machine Learning in Python: Main Devel-
opments and Technology Trends in Data Science, Machine Learning, and Artificial Intelli-
gence. Information, 11(4), 193. Retrieved 2023-03-27, from [Link]
2078-2489/11/4/193 doi: 10.3390/info11040193
Ripley, B. D. (1996). Pattern Recognition and Neural Networks. Cambridge University Press.
doi: 10.1017/CBO9780511812651
Rozemberczki, B. (2023). PyTorch Geometric Temporal Documentation. Retrieved 2023-03-27,
from [Link]
[Link]
Sanchez-Lengeling, B., Reif, E., Pearce, A., & Wiltschko, A. B. (2021, September). A Gentle
Introduction to Graph Neural Networks. Distill, 6(9), e33. Retrieved 2023-02-04, from
[Link] doi: 10.23915/distill.00033
Schörner, P., Hubschneider, C., Härtl, J., Polley, R., & Zöllner, J. M. (2019, October). Grid-
Based Micro Traffic Prediction using Fully Convolutional Networks. In 2019 IEEE Intelli-
gent Transportation Systems Conference (ITSC) (pp. 4540–4547). doi: 10.1109/ITSC.2019
.8917263
Sergios Karagiannakos, A. S. (2021, Sep). Best Graph Neural Network architectures: GCN,
GAT, MPNN and more. Retrieved 2022-11-03, from [Link]
-architectures/
Shahid, N., Shah, M. A., Khan, A., Maple, C., & Jeon, G. (2021, September). Towards greener
smart cities and road traffic forecasting using air pollution data. Sustainable Cities and
Society, 72, 103062. doi: 10.1016/[Link].2021.103062
References 82

Shi, R., & Du, L. (2022, October). Multi-Section Traffic Flow Prediction Based on MLR-LSTM
Neural Network. Sensors (Basel, Switzerland), 22(19), 7517. doi: 10.3390/s22197517
Shuman, D. I., Narang, S. K., Frossard, P., Ortega, A., & Vandergheynst, P. (2013, May). The
Emerging Field of Signal Processing on Graphs: Extending High-Dimensional Data Anal-
ysis to Networks and Other Irregular Domains. IEEE Signal Processing Magazine, 30(3),
83–98. doi: 10.1109/MSP.2012.2235192
Shuvo, M. A. R., Zubair, M., Purnota, A. T., Hossain, S., & Hossain, M. I. (2021, January).
Traffic Forecasting using Time-Series Analysis. In 2021 6th International Conference on
Inventive Computation Technologies (ICICT) (pp. 269–274). doi: 10.1109/ICICT50816
.2021.9358682
Smith, B. L., Williams, B. M., & Keith Oswald, R. (2002, August). Comparison of parametric
and nonparametric models for traffic flow forecasting. Transportation Research Part C:
Emerging Technologies, 10(4), 303–321. doi: 10.1016/S0968-090X(02)00009-8
State of California. (2023). PeMS Data Source | Caltrans. [Link]
traffic-operations/mpr/pems-source.
Sun, H., Liu, H. X., Xiao, H., He, R. R., & Ran, B. (2003). Use of local linear regression model
for short-term traffic forecasting. Transportation Research Record, 1836(1), 143-150. doi:
10.3141/1836-18
Tedjopurnomo, D. A., Bao, Z., Zheng, B., Choudhury, F. M., & Qin, A. K. (2022). A survey
on modern deep neural network for traffic prediction: Trends, methods and challenges.
IEEE Transactions on Knowledge and Data Engineering, 34(4), 1544-1561. doi: 10.1109/
TKDE.2020.3001195
United Nations, U. (2017, September). Rapid urbanisation: opportunities and challenges
to improve the well-being of societies. [Link]
rapid-urbanisation-opportunities-and-challenges-improve-well
-being-societies.
V, R., & S, G. V. (2022, July). Hybrid Time-Series Forecasting Models for Traffic Flow Prediction.
Promet, 34(4), 537–549. Retrieved 2023-01-15, from [Link]
[Link]/PROMTT/article/view/15 doi: 10.7307/ptt.v34i4.3998
van den Oord, A., Dieleman, S., Zen, H., Simonyan, K., Vinyals, O., Graves, A., . . . Kavukcuoglu,
K. (2016). Wavenet: A generative model for raw audio. ArXiv, abs/1609.03499.
Velickovic, P., Cucurull, G., Casanova, A., Romero, A., Lio’, P., & Bengio, Y. (2017). Graph
attention networks (Vol. abs/1710.10903).
Vlahogianni, E., Golias, J., & Karlaftis, M. (2004, 09). Short-term traffic forecasting:
Overview of objectives and methods. Transport Reviews, 24, 533-557. doi: 10.1080/
0144164042000195072
Vlahogianni, E. I., Karlaftis, M. G., & Golias, J. C. (2014, June). Short-term traffic forecast-
ing: Where we are and where we’re going. Transportation Research Part C: Emerging
Technologies, 43, 3–19. doi: 10.1016/[Link].2014.01.005
Wang, Y., Li, L., & Xu, X. (2017, 10). A piecewise hybrid of arima and svms for short-term traffic
References 83

flow prediction. In Neural information processing (p. 493-502). doi: 10.1007/978-3-319

-70139-4_50
Weatherbit. (2023a). Weatherbit - Free Weather API. Retrieved 2023-02-21, from https://
[Link]/
Weatherbit. (2023b). Weatherbit | Sub-Hourly Weather History API Documentation. https://
[Link]/api/weather-history-subhourly. Retrieved 2023-02-17,
from [Link]
WHO, W. H. O. (2022). Billions of people still breathe unhealthy air: new WHO
data. Retrieved 2022-11-03, from [Link]
-billions-of-people-still-breathe-unhealthy-air-new-who-data
Wu, Y., & Tan, H. (2016, December). Short-term traffic flow forecasting with spatial-temporal
correlation in a hybrid deep learning framework. arXiv. doi: 10.48550/arXiv.1612.01022
Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., & Yu, P. S. (2021a). A comprehensive survey
on graph neural networks. IEEE Transactions on Neural Networks and Learning Systems,
32(1), 4-24. doi: 10.1109/TNNLS.2020.2978386
Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., & Yu, P. S. (2021b, January). A Comprehensive
Survey on Graph Neural Networks. IEEE Transactions on Neural Networks and Learning
Systems, 32(1), 4–24. doi: 10.1109/TNNLS.2020.2978386
Wu, Z., Pan, S., Long, G., Jiang, J., & Zhang, C. (2019). Graph wavenet for deep spatial-temporal
graph modeling. In International joint conference on artificial intelligence.
Xie, Q., Guo, T., Chen, Y., Xiao, Y., Wang, X., & Zhao, B. Y. (2019). "how do urban incidents
affect traffic speed?" a deep graph convolutional network for incident-driven traffic speed
prediction. ArXiv, abs/1912.01242.
Yan, H., Fu, L., Qi, Y., Yu, D.-J., & Ye, Q. (2022, August). Robust ensemble method for short-
term traffic flow prediction. Future Generation Computer Systems, 133, 395–410. doi:
10.1016/[Link].2022.03.034
Yao, H., Tang, X., Wei, H., Zheng, G., Yu, Y., & Li, Z. (2018, March). Modeling Spatial-
Temporal Dynamics for Traffic Prediction. ArXiv. Retrieved 2023-01-18, from https://
[Link]/paper/Modeling-Spatial-Temporal-Dynamics
-for-Traffic-Yao-Tang/dc5d1b5a1165aab6d042f48ea7b068a90a900d68
Ye, J., Xue, S., & Jiang, A. (2022, June). Attention-based spatio-temporal graph convolutional
network considering external factors for multi-step traffic flow prediction. Digital Commu-
nications and Networks, 8(3), 343–350. doi: 10.1016/[Link].2021.09.007
Ye, J., Zhao, J., Ye, K., & Xu, C. (2022, May). How to build a graph-based deep learning
architecture in traffic domain: A survey. IEEE Transactions on Intelligent Transportation
Systems, 23(5), 3904-3924. doi: 10.1109/TITS.2020.3043250
Yin, X., Wu, G., Wei, J., Shen, Y., Qi, H., & Yin, B. (2022, June). Deep Learning on Traffic
Prediction: Methods, Analysis, and Future Directions. IEEE Transactions on Intelligent
Transportation Systems, 23(6), 4927–4943. doi: 10.1109/TITS.2021.3054840
Yu, B., Yin, H., & Zhu, Z. (2018, July). Spatio-Temporal Graph Convolutional Networks: A
References 84

Deep Learning Framework for Traffic Forecasting. In Proceedings of the Twenty-Seventh

International Joint Conference on Artificial Intelligence (pp. 3634–3640). doi: 10.24963/
ijcai.2018/505
Yu, T., Yin, H., & Zhu, Z. (2019). St-unet: A spatio-temporal u-network for graph-structured time
series modeling. ArXiv, abs/1903.05631.
Yuan, H., & Li, G. (2021, March). A Survey of Traffic Prediction: from Spatio-Temporal Data
to Intelligent Transportation. Data Science and Engineering, 6(1), 63–85. doi: 10.1007/
s41019-020-00151-z
Zhang, Q., Chang, J., Meng, G., Xiang, S., & Pan, C. (2020, April). Spatio-Temporal Graph Struc-
ture Learning for Traffic Forecasting. Proceedings of the AAAI Conference on Artificial In-
telligence, 34(01), 1177–1185. Retrieved 2022-10-26, from [Link]
[Link]/AAAI/article/view/5470 doi: 10.1609/aaai.v34i01.5470
Zhao, L., Song, Y., Zhang, C., Liu, Y., Wang, P., Lin, T., . . . Li, H. (2020, September). T-GCN:
A Temporal Graph Convolutional Network for Traffic Prediction. IEEE Transactions on
Intelligent Transportation Systems, 21(9), 3848–3858. doi: 10.1109/TITS.2019.2935152
Zhao, Y., Zhang, H., An, L., & Liu, Q. (2018, December). Improving the approaches of traffic
demand forecasting in the big data era. Cities, 82, 19–26. doi: 10.1016/[Link].2018.04.015
Zheng, C., Fan, X., Wang, C., & Qi, J. (2019). Gman: A graph multi-attention network for traffic
prediction. ArXiv, abs/1911.08415.
Zhong, W., Suo, Q., Jia, X., Zhang, A., & Su, L. (2021, July). Heterogeneous Spatio-Temporal
Graph Convolution Network for Traffic Forecasting with Missing Values. In 2021 IEEE
41st International Conference on Distributed Computing Systems (ICDCS) (pp. 707–717).
DC, USA: IEEE. doi: 10.1109/ICDCS51616.2021.00073
Zhou, K., Liu, Z., Duan, K., & Hu, X. (2022). Graph Neural Networks: AutoML. In L. Wu, P. Cui,
J. Pei, & L. Zhao (Eds.), Graph Neural Networks: Foundations, Frontiers, and Applications
(pp. 371–389). Singapore: Springer Nature. doi: 10.1007/978-981-16-6054-2_17
Zhou, X., Shen, Y., & Huang, L. (2019). Revisiting flow information for traffic prediction.
Zhu, J., Song, Y., Zhao, L., & Li, H. (2020, June). A3T-GCN: Attention Temporal Graph Convo-
lutional Network for Traffic Forecasting. arXiv. doi: 10.48550/arXiv.2006.11583