Graf Representation Learning

Content

Objective

This course aims to teach the underlying theory and techniques for transforming graphs—used for data modeling in various fields—into numerical vectors through next-generation representation learning methods. It covers the subject in a broad spectrum, ranging from traditional spectral methods to contemporary Graph Neural Network (GNN) techniques. The primary objective is to equip students with the necessary tools to construct complex systems logic for data analysis and to select the appropriate representation learning technique to solve the problems they encounter.

Course Content

Introduction and Foundations of Graph Theory
Traditional Graph Statistics and Kernel Methods
Neighborhood Overlap and Spectral Methods
Shallow Node Embeddings and Encoder-Decoder Framework
Random Walk Methods and Knowledge Graphs
Graph Neural Networks (GNN) and Message Passing
Aggregation and Update Methods in GNN Architectures
Midterm Exam
Graph Pooling and Relation Prediction Applications
Efficiency in GNN Applications and Node Sampling
Spectral Graph Convolutions and Theoretical Motivations
GNN Capacity and Graph Isomorphism
Traditional and Deep Generative Graph Models
Project Presentation

Course Learning Outcomes

- Ability to handle a complex network as a graph
- Ability to transform a graph into numerical vectors suitable for the studied problem

Teaching and Learning Methods

Lecture, brainstorming, discussion

References

https://www.cs.mcgill.ca/~wlh/grl_book/files/GRL_Book.pdf
http://web.stanford.edu/class/cs224w/

Weekly Contents

Assessment System

Relation of Proficiency

ECTS