GSEK: Graph Structured Neural Networks with Energy Graph Kernel

Graph neural networks (GNNs) have become an increasingly popular framework for describing graph-structured data using deep learning. However, traditional GNNs may not accurately map features for graphs that contain noise or missing information. In order to address this issue, we propose a new method called Graph Structured Neural Networks with Energy Graph Kernel (GSEK) for classification. GSEK defines energy properties for nodes and extracts node subgraphs using maximum energy random wandering to ensure the centrality of the subgraphs' structure. It then uses Wasserstein Weisfeiler-Lehman (WWL) graph kernel to calculate isomorphism coefficients for subgraphs of different node pairs, which are used to define weight matrices and achieve graph reconstruction. Unlike traditional methods that solely focus on node features, GSEK places greater emphasis on the local structure of the graph. Our experiments on six publicly available graph classification datasets demonstrate that our proposed model outperforms other benchmark models.