A new Graph Pooling Method based on Topology and Attribute Features in Graph Neural Networks

Graph neural networks have shown remarkable performance in graph-structured data. However, current research mainly focuses on designing graph convolution operations, while the pooling operation, which is crucial for graph classification tasks, has received insufficient attention. Existing graph pooling methods suffer from the problem of losing graph topology information, resulting in insufficient node feature information mining. Moreover, they only utilize first-order statistics and fail to utilize second-order statistics. In this work, we propose a novel pooling method consisting of two parts. First, we select important nodes based on both attribute and topology features, and then use these nodes to form a pooling subgraph that preserves rich features. Second, we implement a second-order pool to retain higher-order features, which can encode the feature correlation and topology information of all nodes. Our proposed pooling module can be integrated with GCN layers to form a hierarchical pooling structure for graph classification tasks. Experimental results on benchmark datasets demonstrate the superiority of our method.