Visualize GNN message passing layers on a small graph. Each node aggregates neighbor features, demonstrating the Weisfeiler-Lehman hierarchy of expressiveness. Watch features propagate and colors evolve.
Message passing GNN: h_v^{(l+1)} = σ(W·AGG({h_u^{(l)}: u∈N(v)∪{v}})). GCN (Kipf 2017): A_hat·H·W with renormalization. Expressiveness bounded by 1-WL color refinement — cannot distinguish some non-isomorphic graphs. GIN (Xu 2019): sum aggregation with MLP achieves WL power. Graph Transformer uses attention for permutation-equivariant global pooling.