Ego Attention Network: Local Awareness Boosts Graph Attention for Graph Representation Learning

Graph neural networks are now widely studied due to the success of graph representation learning ability, and graph attention networks (GATs) is one of the most popular state-of-the-art. By adapting attention mechanisms to the graph domain, GATs derive their expressive power through a learnable process of aggregating neighboring node features based on weights. Nonetheless, the inherent attention mechanism within GATs predominantly hinges upon individual nodes without taking into account their topological structure, which may lead to feature overfitting and sensitivity to structural noise. To tackle these constraints, we propose a novel graph attention network with local awareness: Ego Attention Network (EAT). EAT replaces node-based attention with local-based attention to exploit both node features and node topology in a local region (egonet specifically). With the egonet embedding, EAT is able to construct smooth and neighbor isomorphism discriminative attention scores. Furthermore, the local structural information is propagated to the updated representation rather than just used to compute the weighting coefficients. Extensive validations of our proposed EAT have been conducted across a gamut of widely recognized graph benchmark datasets. The empirical results demonstrate a marked enhancement in classification performance and robustness against structural noise. The source code is publicly available at:https://github.com/Leejiaxing/EAT