GraphSAGE / GCN / GAT
Best for: Scalable graph representation learning
How it works
$$H^{(l+1)}=\sigma\bigl(\tilde D^{-1/2}\tilde A\tilde D^{-1/2}H^{(l)}W^{(l)}\bigr)$$Concrete GNN variants. GCN propagates with the normalised adjacency $\tilde A=A+I$: $H^{(l+1)}=\sigma\bigl(\tilde D^{-1/2}\tilde A\tilde D^{-1/2}H^{(l)}W^{(l)}\bigr)$. GraphSAGE samples a fixed neighbour set and concatenates, $h_v=\sigma\bigl(W\cdot\text{CONCAT}(h_v,\text{AGG}(\{h_u:u\in N(v)\}))\bigr)$. GAT replaces aggregation with learned attention $\alpha_{ij}=\text{softmax}_j\bigl(\text{LeakyReLU}(a^\top[Wh_i\|Wh_j])\bigr)$, weighting different neighbours differently.
Common fields
Social networks · recommender systems · cybersecurity