A talk by Aleksa Gordic | Software - Deep Learning engineer, Microsoft | The AI Epiphany What can you learn about Graph Machine Learning in 2 months? Aleksa Gordic, Machine Learning engineer @ Microsoft and Founder @ The AI Epiphany, shares his journey in the world of Graph Machine Learning. Aleksa started exploring the basics in the world of Graph Machine Learning, and ended up implementing and open sourcing his own Graph Attention Network on PyTorch. In this talk, Aleksa will share the fundamentals of Graph Machine Learning, provide real-world examples, resources, and everything his younger self would be grateful for. Aleksa will also be available to answer questions. What is Graph Machine Learning? Simply put, Graph Machine Learning is a branch of machine learning that deals with graph data. Graphs consist of nodes, that may have feature vectors associated with them, and edges, which again may or may not have feature vectors attached. The applications are endless. Massive-scale recommender systems, particle physics, computational pharmacology / chemistry / biology, traffic prediction, fake news detection, and the list goes on and on.

1. The
magic
world of
Graph ML

2. What are graphs?
Nodes Edges

4. 50 shades of graphs
Simplifying assumption:
• Undirected
• Self-edges

5. What we care about: features
Classic node features example: Cora
• Vocab
• Is word present in this paper/node?
Edge features example: molecules

6. Graph Neural Networks (GNNs) Abstract applications:
• Graph classification/regression
• Node/edge classification/regression

7. Graph ML - applications

8. Graph ML apps

9. Graph ML apps – fundamental science

10. Contrasting Graph ML with CV, NLP, RL

11. Graph ML – high level survey

12. Graph embedding methods
DeepWalk, Node2Vec, etc.

13. Convolution (on the way from CV to Graph ML)

14. Adjacency matrix, Laplacian, signals

15. GNNs – spectral methods
Cons:
• Computationally expensive ~O(n³)
• Inherently transductive
Pros:
• Mathematical notion of convolution
Chebnets:
k-hop, faster

16. Spatial GNNs

17. GNN expressivity (going beyond WL)
One solution: counting subgraphs
Are these graphs the same?
More expressive -> better for your problem? No.

18. Dynamic graphs
• Representing time as a vector
• Temporal neighborhood

19. GAT project

20. GAT – Graph Attention Network project

21. Graph Attention Network (GAT) project

22. Further learning resources for Graph ML:
• https://github.com/gordicaleksa/pytorch-GAT (Graph Attention Net)
• https://www.youtube.com/c/TheAIEpiphany (lots of Graph ML vids)
• https://gordicaleksa.medium.com/how-to-get-started-with-graph-
machine-learning-afa53f6f963a (getting started with Graph ML)