Computer science

Case study: May I recommend the following?

For use in May and November 2023
___________________________________________________________________________

Introduction
You have just ended a video conferencing call with your two friends, Jungmin and Lijing.
They have an idea for an online business venture and they want your help. The business will
be called NextStar. It will provide an application so that users can view the work of artists
who have yet to be discovered.
Artists may include actors, singers, screenwriters, comedians, painters, sculptors and
filmmakers. In fact, any artist who wants to demonstrate a talent will be able to upload files
to the application. The uploaded content can be rated by all users. Based on these ratings, the
application recommends new content to each user.
Jungmin and Lijing plan to make the NextStar website free to join and believe that they will
be able to make money from advertising once there are enough users. They realize that this
application will eventually require a great deal of storage, so they are looking at cloud-
hosting companies. Once enough content has been added, the application will incorporate a
recommender system.
The following information provides an outline of what has already been researched and
includes some challenges for you to consider

Cloud computing
Hosting applications that utilize data at an enterprise level are widely available and affordable
thanks to cloud computing. Users only pay for the resources they use, so they can start small
and add more resources as they grow. This makes cloud computing ideal for a start-up like
NextStar.
There are a number of cloud deployment models that could be used to host NextStar’s data.
There are also three cloud delivery models: software as a service (SaaS), platform as a
service (PaaS), and infrastructure as a service (IaaS). NextStar intends to use IaaS.

Machine learning
Machine learning is a subfield of artificial intelligence. There are three main types of
machine learning: supervised learning, unsupervised learning and reinforcement learning (see
Figure 1).
 Figure 1: The three main types of machine learning

A supervised learning algorithm uses labelled training data to learn a function that produces
an appropriate output when given new unlabelled data. Typically, supervised learning is used
to classify data or make predictions.
An unsupervised learning algorithm learns patterns from unlabelled data. These algorithms
draw references from observations of the live input data. The system can organize data into
subsets, or clusters, that have not been pre-classified by the programmers.
A reinforcement learning algorithm learns in an interactive environment by trial and error
using feedback from its own actions and experiences. Some recommender systems can be
seen as a type of reinforcement learning because positive behaviour, such as reviewing
content, is rewarded with better recommendations.

Recommender systems
Where there is a huge amount of content available, a recommender system directs a user to
content that they have not seen but may be of interest to them. Recommender systems use
data for content that users have already rated (actual data) to generate predicted preferences
for content that they have not already rated (predicted data).
The majority of recommender systems utilize supervised learning. The use of unsupervised
learning and reinforcement learning is less common.
Recommender systems can use content-based filtering, collaborative filtering, or a
combination of both. Hybrid recommender systems combine several machine learning
algorithms. This was demonstrated on 21 September 2009, when BellKor’s Pragmatic Chaos
team won the Netflix Prize and USD 1000000 for the best collaborative filtering movie
recommender system. This recommender system combined 107 different algorithms in a
hybrid model that outperformed Netflix’s own algorithm’s root-mean-square error (RMSE)
score by 10.06%.
Content-based filtering
Content-based filtering, sometimes called item-item filtering, focuses on an item’s attributes
rather than using user interactions and feedback. The content-based approach is one of user-
specific classification, in which the classifier learns the user’s likes and dislikes based on an
item’s attributes.
Since NextStar’s recommender system will contain video clips, attributes might include
genre, release date, artist, language, gender, and age. For example, if a user rates stand-up
comedy clips highly, the system is likely to recommend more comedy clips to them (see
Figure 2).
Figure 2: An example of content-based filtering

Collaborative filtering
With collaborative filtering, the recommendations for each user are generated by using the
rating information from other users and items. The core assumption is that users who have
agreed in the past tend to agree in the future. So, if two users scored content similarly, other
content rated highly by one user is likely to be enjoyed by the other user. Thus, that content
can be recommended to the second user (see Figure 3). One of the limitations of
collaborative filtering recommender systems is popularity bias, where popular content is
recommended too frequently.
Figure 3: An example of collaborative filtering
Collaborative filtering can use different algorithms to recommend new content. Two types of
algorithm that can be used are k-nearest neighbour (k-NN) and matrix factorization.
K-nearest neighbour (k-NN)
The k-NN algorithm uses feature similarity to predict the values of any new or missing data.
This means that new data points are assigned a value based on how closely they resemble
other data points in the training set.
The k-NN algorithm makes its predictions based on the nearest neighbours. The “k” aspect of
this algorithm represents the number of neighbours and is simply a hyperparameter that can
be adjusted using a trial-and-error approach.
Matrix factorization
Matrix factorization is an alternative to the k-NN algorithm. The difficulty with using
standard matrix factorization approaches for recommender systems is that the dataset is not
complete.
To overcome this limitation, values need to be estimated for the smaller matrices using an
iterative algorithm.
In Figure 4, the user–item interaction matrix represents each user’s rating (rows) of each
content item (columns). User 1, represented by US1, has rated the first three items but has not
rated item 4 or item 5, represented by IT4 and IT5.
Figure 4: User–item interaction matrix

Matrix factorization works by decomposing the large user–item interaction matrix into two
smaller matrices—an item–feature matrix and a user–feature matrix—to capture the most
important features required for learning. If the values in the item–feature matrix and the user–
feature matrix are changed, the corresponding values in the user–item interaction matrix will
also change (see Figure 5).
 Figure 5: Matrix factorization
Item–feature matrix

The matrices are tuned by generating predicted preferences for content where actual
preference data already exists. Once the prediction values approach the actual rating, the
assumption is that the matrices will be able to effectively predict preferences for which no
actual data exists.
A process called stochastic gradient descent uses a cost function to adjust each cell by
making a small change to the item–feature and user–feature matrices. For example, in Figure
4, the value in the US1 and IT1 intersecting cell is 3, but in Figure 5 this value is 2.18. So,
the error for this cell is (3 - 2.18)2, or 0.6724.
Training recommender systems
A recommender system can be evaluated using train/test splits. The ratings data is split into a
training set and a testing set. A commonly used split is when 80% of the data is assigned to
the training set and the other 20% to the testing set.
A recommender system learns the relationships between items and the relationships between
users. Once trained, it makes predictions about how a user might rate an item that they
haven’t rated yet.
A common problem of training a machine learning algorithm is overfitting, where the model
fits too closely to the training dataset. When the model trains for too long on the training data,
or when the model is too complex, it can start to learn the irrelevant features within the
dataset. Consequently, the model fails to generalize effectively against new data.
Evaluating recommender systems
Recommender system accuracy can be evaluated through two different measures: mean
absolute error (MAE) and root-mean-square error (RMSE). These measures give an
indication of how the recommender systems perform on training/test data.
However, the effectiveness of a recommender system is not fully known until it has been
used by the public. A recommender system is not performing well if it fails to recommend
content the user would like or recommends content that they do not like.
Precision and recall are performance metrics used on live data. Precision is a measure of
exactness, the fraction of relevant instances among the retrieved instances. Recall is a
measure of completeness. The F-measure provides a single score that balances the concerns
of precision and recall.
The way that recommendations are displayed to users is also important. A list might be
sufficient, or it may be possible to select recommended content by groups or subgroups.
These groups might be organized by genre, gender, age or any number of possible categories.

Social and ethical concerns

When building a model from users’ behaviour, two types of behavioural data can be used:
explicit data and implicit data.
Explicit behavioural data refers to data gathered from users’ submitted data, such as when a
user rates a video clip, enters their preference, or searches for an item. Users may believe this
is the only data that is used to make recommendations.
Implicit behavioural data refers to data that the user is not aware is being collected. This
might include click data, purchase data, or even the use of a key logger.
The quality of user data is critical to the success of the NextStar project, but there are ethical
concerns about the collection, storage and use of behavioural data. NextStar also needs to
consider its users’ right to anonymity and right to privacy.

Challenges faced
To help your friends with their new business venture, there are a number of challenges that
you need to research:
 Understanding the similarities and differences between supervised learning,
unsupervised learning and reinforcement learning.
 Understanding how the k-NN algorithm and matrix factorization can be used within
recommender systems.
 Understanding how to train, test and evaluate a recommender system.
 Comparing content-based filtering and collaborative filtering recommender systems.
 Understanding the ethical concerns linked to the collection, storage and use of users’
behavioural data.
Candidates are not required to know the mathematical equations relating to
recommender systems.
Additional terminology
Behavioural data
Cloud delivery models:
Infrastructure as a service (IaaS)
Platform as a service (PaaS)
Software as a service (SaaS)
Cloud deployment models
Collaborative filtering
Content-based filtering
Cost function
F-measure
Hyperparameter
K-nearest neighbour (k-NN) algorithm
Matrix factorization
Mean absolute error (MAE)
Overfitting
Popularity bias
Precision
Recall
Reinforcement learning
Right to anonymity
Right to privacy
Root-mean-square error (RMSE)
Stochastic gradient descent
Training data

Some companies, products, or individuals named in this case study are fictitious and
any similarities with actual entities are purely coincidental.

Disclaimer:
Content used in IB assessments is taken from authentic, third-party sources. The views
expressed within them belong to their individual authors and/or publishers and do not
necessarily reflect the views of the IB.
References:
Figure 1 Jones, M. T., 2017. Models for machine learning. [online] Available at:
[Link] [Accessed 15 October 2021]. source adapted.