graph attention networks
**Graph Attention Networks (GAT)** are **neural networks that use attention mechanisms to weight neighbor importance in graphs** — learning which connected nodes matter most for each node's representation, achieving state-of-the-art results on graph tasks.
**What Are GATs?**
- **Type**: Graph Neural Network with attention mechanism.
- **Innovation**: Learn importance weights for each neighbor.
- **Contrast**: GCN treats all neighbors equally, GAT weighs them.
- **Output**: Node embeddings incorporating weighted neighborhood.
- **Paper**: Veličković et al., 2018.
**Why GATs Matter**
- **Adaptive**: Learn which neighbors are important per-node.
- **Interpretable**: Attention weights show reasoning.
- **Flexible**: No fixed aggregation (unlike GCN averaging).
- **State-of-the-Art**: Top performance on citation, protein networks.
- **Inductive**: Generalizes to unseen nodes.
**How GAT Works**
1. **Compute Attention**: Score importance of each neighbor.
2. **Normalize**: Softmax across neighbors.
3. **Aggregate**: Weighted sum of neighbor features.
4. **Multi-Head**: Multiple attention heads, concatenate results.
**Attention Mechanism**
```
α_ij = softmax(LeakyReLU(a · [Wh_i || Wh_j]))
h'_i = σ(Σ α_ij · Wh_j)
```
**Applications**
Citation networks, protein-protein interaction, social networks, recommendation systems, molecule property prediction.
GAT brings **attention to graph learning** — enabling adaptive, interpretable node representations.