Graph Neural Network
A GNN updates each node from its neighbors. Message passing lets the model use relational structure — markets, molecules, citation graphs — instead of forcing a grid.
Definition
Message-passing GNNs compute h_v^{(k+1)} = UPDATE(h_v^{(k)}, AGGREGATE({h_u^{(k)} : u ∈ N(v)})). GCN (Kipf & Welling) is a spectral-inspired average; GAT learns attention over neighbors. Depth is limited by over-smoothing: too many hops make nodes look alike.
Why it matters
Whenever the object is a graph — counterparties, supply chains, limit-order events, knowledge graphs — a GNN is the architecture that respects the domain. Flattening to a table throws the edges away.
Case
A GCN on a citation graph classifies papers from neighbor topics. The same idea on a correlation graph of equities is only as good as the graph you built — a fully-connected correlation matrix is not structure, it is a dense blob.
How to read it
Specify the graph construction. If edges are thresholded correlations recomputed in-sample, the GNN may be leaking the label through the adjacency matrix.
Key papers
Landmark references with real publisher or arXiv links — not token-overlap noise.
- Semi-Supervised Classification with Graph Convolutional NetworksKipf, Welling · 2017 · ICLRGCN.
- Graph Attention NetworksVeličković et al. · 2018 · ICLRGAT.