Abstract
Graph Neural Networks (GNNs) serve as a powerful framework for representation learning on graph-structured data, capturing the information of nodes by recursively aggregating and transforming the neighboring nodes' representations. Topology in graph plays an important role in learning graph representations and impacts the performance of GNNs. However, current methods fail to adequately integrate topological information into graph representation learning. To better leverage topological information and enhance representation capabilities, we propose the Graph Topology Attention Networks (GTAT). Specifically, GTAT first extracts topology features from the graph's structure and encodes them into topology representations. Then, the representations of node and topology are fed into cross attention GNN layers for interaction. This integration allows the model to dynamically adjust the influence of node features and topological information, thus improving the expressiveness of nodes. Experimental results on various graph benchmark datasets demonstrate GTAT outperforms recent state-of-the-art methods. Further analysis reveals GTAT's capability to mitigate the over-smoothing issue, and its increased robustness against noisy data.