• Start the environment with "conda activate p_ppnp"

• To train all models you can execute the ./run_scripts/run.sh file (The MCD versions and the DE version are not trained as they use the pretrained corresponding standard model).

• To evaluate all models you can execute the ./run_scripts/load.sh file.

• The default dataset is Cora. Change "cora" to "citeseer" in run.sh and load.sh in order to train and evaluate on Citeseer.

• pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116
• pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.12.0+cu116.html

• Tested under Ubuntu/WSL

• Install CUDA and cuDNN for GPU support.

• Setup the conda environment: conda env create -f p_ppnp_environment.yml python==3.9 --force

• Log in to WandB and change the project and entity for wandb.init in /pgnn/logger/logger.py

• Permission to execute the files run.sh and load.sh might need to be given ("chmod +x run.sh", "chmod +x load.sh")

• If you use the networkx_to_sparsegraph method for importing other datasets you will additionally need NetworkX.

As the basis for this project we relied on the original PPNP implementation (basic structure, data, preprocessing, training, PPNP model, etc.): Johannes Gasteiger, Aleksandar Bojchevski and Stephan Günnemann. "Predict then Propagate: Graph Neural Networks meet Personalized PageRank." ICLR 2019. The github repository for this project: https://github.com/gasteigerjo/ppnp

The OOD Detection and the GPN model originate from: M. Stadler, B. Charpentier, S. Geisler, D. Zügner, and S. Günnemann. "Graph Posterior Network: Bayesian Predictive Uncertainty for Node Classification." 2021. doi: 10.48550/ARXIV.2110.14012. The github repository for this project: https://github.com/stadlmax/Graph-Posterior-Network

For the PyTorch GCN implementation, we utilized code from the original TensorFlow implementation: T. N. Kipf and M. Welling. “Semi-Supervised Classification with Graph Convolutional Networks.” In: CoRR abs/1609.02907 (2016). arXiv: 1609.02907. The github repository for this project: https://github.com/tkipf/gcn

The GAT implementation in PyTorch was provided by: A. Gordić. pytorch-GAT. https://github.com/gordicaleksa/pytorch-GAT. 2020. The github repository for this project: https://github.com/gordicaleksa/pytorch-GAT

The according licenses for the listed projects are provided in ./licenses

In the data folder you can find several datasets. If you want to use other (external) datasets, you can e.g. use the networkx_to_sparsegraph method in ppnp.data.io for converting NetworkX graphs to our SparseGraph format.

The Cora-ML graph was extracted by Aleksandar Bojchevski, and Stephan Günnemann. "Deep gaussian embedding of attributed graphs: Unsupervised inductive learning via ranking." ICLR 2018,
while the raw data was originally published by Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore. "Automating the construction of internet portals with machine learning." Information Retrieval, 3(2):127–163, 2000.

The Citeseer graph was originally published by Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Gallagher, and Tina Eliassi-Rad. "Collective Classification in Network Data." AI Magazine, 29(3):93–106, 2008.