Node feature propagation, aggregation, and classification via GNN rounds
0
■ Node color = class
Brightness = feature[0]
──── Edge (message)
Ring = embedding dim 0
GNN Message Passing: h⁽ˡ⁺¹⁾ᵥ = σ(W·AGG({h⁽ˡ⁾ᵤ : u ∈ N(v) ∪ {v}})).
Features start as initial node attributes. Each round aggregates neighbor embeddings (mean/sum/max), applies a learned transform (simulated here as random projection), then ReLU.
After K rounds, nodes with same color should cluster in embedding space — showing how GNNs perform graph-structure-aware classification.