Graph Attention Network: Node Classification

Attention Heads

Layer Parameters

Softmax temperature T 1.00

Feature noise σ 0.60

Readouts

Classification accuracy

—

Graph edges / avg degree

—

Class A Class B Class C

Ring = true label · Sphere fill = GAT prediction

Focused Node

Click a node to inspect its incoming attention weights α_ij.

How it works

A Graph Attention Network layer learns, for every node i, how much to trust each neighbor j when building a new representation. For a shared weight matrix W and attention vector a per head:

e_ij = LeakyReLU( a·[ W h_i || W h_j ] )
α_ij = softmax_j( e_ij / T )   over j in N(i) ∪ {i}
h_i' = ELU( Σ_j α_ij · W h_j )

softmax normalizes the raw scores e_ij over each node's own neighborhood, so α_ij sums to 1 across a node's neighbors — this is exactly the "attention" in graph attention. Multiple independent heads (different W, a) are computed in parallel and concatenated, then a linear readout maps the aggregated features to class logits, softmaxed into a predicted class.

Attention heads — more heads give the layer more independent ways to weigh neighbors before concatenation.
Temperature T — divides e_ij before softmax; low T sharpens attention onto the single best-matching neighbor, high T spreads it near-uniformly.
Feature noise σ — perturbs each node's input features around its true class center, degrading how well attention (and the final classifier) can separate classes.
Click a node — highlights its edges by α_ij and reports the entropy of its attention distribution (low entropy = focused on one neighbor, high = spread out).

This mirrors the original GAT architecture (Veličković et al., 2018) used for semi-supervised node classification on graphs like citation networks, where a node's class is inferred from a learned weighting of its neighbors rather than a fixed average.

Drag to orbit · click a node to inspect attention

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.