Graph Attention Network

The problem, from first principles

Graph Attention Network (GAT) uses attention mechanisms to aggregate neighbor information. Each node computes attention weights for its neighbors, giving more weight to important neighbors.

Where alpha_ij are attention weights computed from node features.

From theory to code

Implement:

gat_attention_weights(features, adj)

Computes attention weights for each edge in the graph.

Constraints

• features shape: (num_nodes, feature_dim).
• adj shape: (num_nodes, num_nodes), binary adjacency.
• Return shape: (num_nodes, num_nodes), attention weights.