A Spanning-tree-based Method for Accelerating Graph Classification

Abstract

Graph neural network (GNN) shows strong processing ability on graph model data. Compared with node-level tasks, the GNN-based graph pooling model for graph-level tasks pays more attention to the overall features of the graph and is often used for protein function prediction, friend network classification, and other practical problems. Recently, many researchers have introduced graph structure information into the graph pooling model to improve the model's accuracy. However, there is still a problem with the long training period, and the influence of structure information on node features is not considered in graph convolution. Therefore, this paper proposes a graph-level feature learning acceleration method based on a spanning tree. This method can construct a spanning tree containing the overall structure information on the undirected graph according to the greedy strategy, called Graph-units, and use the graph attention network to learn the node features containing the graph structure information in the graph convolution process, then input the node features into the graph pooling process to learn the graph level features. Finally, the model can achieve the highest accuracy with fewer iterations. This method is applied to multiple graph pooling models on multiple graph classification task datasets, and an experimental comparison is made to verify the effectiveness of the proposed method.