Graph Neural Networks

Best for: Learning from graph structure

How it works

GNNs learn node representations by neighbourhood message passing. Each layer updates a node by aggregating its neighbours’ features, $h_v^{(l+1)}=\sigma\bigl(W^{(l)}\cdot\text{AGG}(\{h_u^{(l)}:u\in N(v)\})+b\bigr)$, where AGG is a permutation-invariant function (sum, mean, max, or attention). After $L$ layers the embedding $h_v^{(L)}$ encodes the $L$-hop neighbourhood, supporting node classification, link prediction, and graph-level prediction via pooling.

Common fields

Molecules · fraud · knowledge graphs