Graph Neural Networks
Learn from data shaped like networks: node embeddings, message passing, GCN, GAT, GraphSAGE, link prediction, and real uses.
Before this: Deep Learning. Linear algebra helps.
-
14 min
Step 1Graph neural networks from the ground up
Why graphs need their own kind of network.
Graph Neural Networks - a perspective from the ground upAlex Foo
-
16 min
Step 2Node embeddings and node2vec
Use examples/node2vec.py to turn random walks into node embeddings.
Graph Embeddings (node2vec) explained - How nodes get mapped to vectorsPhilipp Brunenberg
pyg-team/pytorch_geometric★ 24.1k -
11 min
Step 3Message passing
The loop every GNN layer repeats: gather, combine, update.
Simple Message Passing on GraphsWelcomeAIOverlords
-
9 min
Step 4Graph convolutional networks
The simplest popular GNN layer, with the intuition.
Graph Convolutional Networks (GCNs) made simpleWelcomeAIOverlords
-
15 min
Step 5Graph attention networks
Use examples/gat.py to train a network that weights its neighbors.
Understanding Graph Attention NetworksDeepFindr
pyg-team/pytorch_geometric★ 24.1k -
13 min
Step 6GraphSAGE
Sample neighbors to train on graphs too big for memory.
GraphSAGE: Inductive Representation Learning on Large Graphs (Graph ML Research Paper Walkthrough)TechViz - The Data Science Guy
-
3 min
Step 7What GNNs can and cannot tell apart
The Weisfeiler-Lehman test as a limit on expressiveness.
Weisfeiler-Lehman WL Test for Graph Isomorphism explained visually & Message Passing NNs 2022Discover AI
-
1h 14m
Step 8Hands-on with PyTorch Geometric
Build and train a graph neural network with PyTorch Geometric.
Graph Neural Networks (GNN) using Pytorch Geometric | Stanford UniversityLindsey AI
pyg-team/pytorch_geometric★ 24.1k -
17 min
Step 9Link prediction
Score candidate links using features computed from the graph.
Stanford CS224W: ML with Graphs | 2021 | Lecture 2.2 - Traditional Feature-based Methods: LinkStanford Online
-
13 min
Step 10Recommenders with GNNs
Treat users and items as a bipartite graph.
Recommender Systems using Graph Neural NetworksDeepFindr
-
19 min
Step 11GNNs that generate molecules
Atoms as nodes, bonds as edges: a GNN proposes new molecules.
Discovering New Molecules Using Graph Neural Networks by Rocío MercadoGAIA
Build it
Watching isn't learning. Pick one and put it on GitHub.
- Classify nodes in Cora with GCN, GAT, and GraphSAGE, and compare the three.
- Predict missing links on a citation graph and report ROC-AUC against a node2vec baseline.
- Train a GNN on the MUTAG molecule dataset in PyTorch Geometric to classify whole graphs, and compare it with a simple baseline.