BK-MSSNN: BI-Kernel Based Multi-Scale Subgraph Neural Networks

Graph Neural Networks (GNNs) have achieved remarkable success in handling graph-related tasks. Most GNN frameworks are based on Message Passing Neural Networks (MPNNs). However, subgraph structures are often complex and multi-scale, and the performance of MPNN in graph isomorphism discrimination tasks may have certain limitations. To address this issue, this paper investigates the semantic associations between subgraphs of different scales and proposes a novel Bi-Kernel Multi-Scale Subgraph Neural Network (BK-MSSNN) to represent more finegrained subgraph semantic information. Specifically, we sample global subgraphs and local subgraphs of different scales, and integrate two graph kernel methods, namely the RW kernel and the Weisfeiler-Lehman kernel, to construct multiple associative relationships between subgraphs for updating node embeddings. Experimental results on five public datasets demonstrate that our proposed model outperforms other benchmark models in graph classification tasks.