Graph Attention Networks (GAT)

Applied masked self-attention to graphs, assigning dynamic, anisotropic attention weights to neighboring nodes during message passing.

Universal graph neural architecture for inductive prediction on protein-protein interaction networks and fraud detection graphs.

Assumptions

• Edges carry non-uniform semantic importance that can be learned end-to-end via multi-head self-attention

Limitations

• Original GAT computes static attention where the ranking of neighbors is query-independent (fixed by Brody et al. in GATv2)