GAT (Graph Attention Networks)

PyTorch implementation of the paper "Graph Attention Networks" (ICLR 2018)

Requirements

• Python 3.6 >
• PyTorch 1.4 >

Results

• GCN : hidden node 64, GAT : hidden node 8, num head 8
• Accuracy on Cora (mean/high/low) : 0.8055/0.8170/0.7930 (GAT), 0.7659/0.7730/0.7530 (GCN)
• Accuracy on Citeseer (mean/high/low) : 0.6333/0.6490/0.6130 (GAT), 0.6161/0.6220/0.6080 (GCN)

Accuracy Curves

Loss Curves

References

[1] Veličković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., & Bengio, Y. (2017). Graph attention networks. arXiv preprint arXiv:1710.10903.

[2] Kipf, T. N., & Welling, M. (2016). Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907.