Introduction to Graph Machine Learning

Abstraction: Graph neural network fundamentals from representations to transformers

Key points:

• Graph tasks span node-level (AlphaFold protein folding uses node property prediction), edge-level (drug side-effect prediction), and graph-level (molecule toxicity classification) — plus subgraph community detection
• GNN core: message passing aggregates neighbor representations each layer; after n layers a node encodes its full n-hop neighborhood
• Oversmoothing problem: too many GNN layers collapse all node representations toward a single vector, preventing scaling unlike Transformers
• Key GNN variants: Graph Convolutional Networks (average neighbors), Graph Attention Networks (learned weights like Transformers), GraphSAGE (sampled multi-hop + max pooling), Graph Isomorphism Networks (MLP on summed neighbors)
• Transformers without positional encodings are permutation-invariant, enabling graph use; TokenGT treats nodes and edges as token sequences; GraphGPS is a framework combining message passing with linear Transformers
• Practical libraries: PyTorch Geometric, Deep Graph Library, NetworkX; benchmarks: OGB (Open Graph Benchmark)

Connections: Hugging Face · Deepmind · Graph Neural Networks · Machine Learning · Transformers · Message Passing

Source: https://huggingface.co/blog/intro-graphml