Substructure Aware Graph Neural Networks

Despite the great achievements of Graph Neural Networks (GNNs) in graph learning, conventional GNNs struggle to break through upper limit expressiveness first-order Weisfeiler-Leman isomorphism test algorithm (1-WL) due consistency propagation paradigm with 1-WL.Based on fact that it is easier distinguish original subgraphs, we propose a novel framework neural network called Substructure Aware (SAGNN) address these issues. We first Cut subgraph which can be obtained from by continuously and selectively removing edges. Then extend random walk encoding return probability rooted node capture structural information use as feature improve GNNs. theoretically prove our more powerful than 1-WL, superior structure perception. Our extensive experiments demonstrate effectiveness framework, achieving state-of-the-art performance variety well-proven tasks, equipped perform flawlessly even 3-WL failed graphs. Specifically, achieves maximum improvement 83% compared base models 32% previous methods.