Multi-scale graph classification with shared graph neural network

Peng Zhou, Zongqian Wu, Guoqiu Wen, Kun Tang, Junbo Ma

Abstract: Graph data is an irregular structural data type that is broadly used in various realistic scenarios to represent the complex interrelationships or topological structures inside the data. As the topological structures of the graphs are usually different from each other, it is difficult to handle the graph classification task. Most existing methods rely on Graph Neural Networks (GNNs) to extract the graph embeddings and produce the classification results based on these graph embeddings. However, these GNN-based methods usually have the over-smoothing issue caused by superimposing GNN to increase the receptive field when extracting high-order local structures. To address this issue, we proposed a novel Multi-Scale Fusion Graph Neural Network (MSFG) in this paper. Through the proposed multi-scale graph coarsen framework and parameter sharing mechanism, the proposed model can efficiently extract high-order structural features of graphs without increasing the number of GNN layers. We conduct experiments on six graph classification datasets and the experimental results show the effectiveness of the proposed MSFG model. Furthermore, multiple ablation experiments prove the validity of each component of the proposed model.