Edge-Gated Graph Attention Network for Graph Classification

Accurately predicting graph-level properties from graph-structured data is a fundamental task in machine learning. However, traditional Graph Attention Networks (GATs) suffer from two key limitations: they often neglect edge attribute information and are prone to gradient vanishing in deep architectures. To address these issues, we propose the Edge-Gated Graph Attention Network (EGGAT), a novel architecture for graph classification. EGGAT introduces a dynamic edge gating mechanism and a triple attention mechanism to effectively capture node-edge-node interactions. Additionally, EGGAT employs residual connections and batch normalization for stable deep training, and integrates FLAG (Free Large-scale Adversarial Augmentation on Graphs) to improve generalization under limited supervision. On the ogbg-molhiv benchmark, EGGAT achieves a test AUC of 0.8066, outperforming lightweight GNNs with similar parameter counts and surpassing the pre-trained Graphormer despite using only 1.2% of its parameters.