Graph Attention Network: Attention Matrix View

This simulator runs a real Graph Attention Network (GAT) layer over a 24-node graph, rendered entirely in 2D. Each node carries a feature vector drawn from one of three class centers; a directed k-nearest-neighbor graph connects it to its most feature-similar peers, and a 2D force (spring) layout — not a flattened 3D projection — spreads the nodes across the canvas. On every parameter change the layer computes LeakyReLU attention scores for each edge, softmax-normalizes them per neighborhood exactly as in the GAT paper, aggregates neighbor features into a new representation, and classifies each node with a linear readout. A second panel renders the full N×N attention-weight matrix as a heatmap — a genuinely 2D-native way to see every node's attention distribution at once, something a single-node 3D inspector can't show. Click any node or matrix row to see the raw attention weights it places on its own neighbors and how concentrated (low-entropy) or diffuse (high-entropy) that attention is.