Improved Hierarchical Architecture for Graph Classification

The existing methods of graph classification often do not consider the information of the graph from multiple perspectives, nor do they dynamically fuse the relevant information scores of nodes, which will lead to incomplete graph structure information. Additionally, after sending the coarse graphs obtained from the pooling layer to the readout layer, the graph representations output by the readout layer is sent to the linear layer directly, which will cause the final generated graph representations containing redundant information. In response to the above issues, two improvements are proposed in this paper: (1) Consider the graph structure from multiple perspectives so that the second-order neighbor information is introduced. Thereby the structural relationship among neighbor nodes can be fully used to strengthen the integrity of graph structure. After that, the structure and feature information is dynamically fused. (2) Analyze the importance of each coarse graph obtained from graph pooling via attention and then designate weights according to their importance. Eventually, the more representative graphs are obtained. The results of the experiments on four different datasets reveal that this fashion can achieve better performance in graph classification tasks.