Graph Neural Network Learning Based on Subgraph Feature

Recently, graph neural networks have achieved success in graph representation learning. However, most conventional graph neural networks adopt a 1-Hop message aggregation strategy, which repeatedly aggregates information from1st-order neighboring nodes. This approach's expressive power is limited by the 1 Weisfeiler-Lehman (1-WL) test. In addition, aggregation method also faces the problem of Long-Range Dependencies, which limits the model to learn different local structural information of a graph and leads to non-discriminatory graph features on some non-isomorphic graphs. Existing work has proposed variants such as k-WL to overcome the discriminative upper limit of1-WL but at a cost. To reduce time costs and improve model resilience to isomorphism-related problems, we propose SubGraphormer that combines subgraph features based on rich and diverse local structural information of the graph. we define subgraphs based on a graph partitioning algorithm and subgraph features according to node-subgraph mappings. Then it integrates subgraph features in the process of message passing aggregation to obtain superior graph feature representations. Comparative and ablation experiments are conducted on a series of benchmark datasets to evaluate SubGraphormer performance, and the results show that combining subgraph features effectively improves performance on various benchmark networks.