Path-aware multi-hop graph towards improving graph learning

Abstract

Graph Neural Networks (GNNs) have achieved state-of-the-art performance in graph-related tasks. Most of them pass messages between direct neighbors and the deeper GNNs can theoretically capture the more global neighborhood information. However, they often suffer from over-smoothing problems when the GNNs’ depth deepens. To eliminate this limitation, we propose a path-aware multi-hop graph framework with hop-level attention (HAPG), which has a larger receptive field than GNNs with the same depth. HAPG generates some multi-path multi-hop graphs by converting the original graph to achieve message passing between multi-hop neighbors in a single layer. HAPG aggregates one-hop and multi-hop neighbors by stacking the original graph and the generated graphs into GNNs. The node embeddings are obtained by combining the multiple outputs of the GNNs with the help of hop-level attention. Comparative experiments with various existing GNNs are conducted on three benchmark datasets, and results show that the HAPG is an effective way for improving these models. Specifically, for semi-supervised node classification tasks, the proposed HAPG-GAT and HAPG-AGNN have achieved state-of-the-art performance.