MVMNET: Graph Classification Pooling Method with Maximum Variance Mapping

Abstract

Graph Neural Networks (GNNs) have been shown to effectively model graph-structured data for tasks such as graph node classification, link prediction, and graph classification. The graph pooling method is an indispensable structure in the graph neural network model. The traditional graph neural network pooling methods all employ downsampling or node aggregating to reduce graph nodes. However, these methods do not fully consider spatial distribution of nodes of different classes of graphs, and making it difficult to distinguish the class of graphs with spatial locations close to each other. To solve such problems, this article proposes a Maximum Variance graph feature Multistructure graph classification method (MVM), which extracts graph information from the perspective of graph nodes feature and graph topology. From the nodes feature perspective, we enlarge the variance between different classes while maintaining the variance between the same class of data. Then the hierarchical graph convolution and pooling are performed from a topological perspective and combined with a CNN readout mechanism to preserve more graph information to obtain a graph-level representation with strong discrimination. Experiments demonstrate that our method outperforms several number of state-of-the-art graph classification methods on multiple publicly available datasets.