Graph Pooling in Graph Neural Networks with Node Feature Correlation

Abstract

Because of the excellent performance of convolutional neural network in computer vision and natural language processing, we extend convolution operation to graph data and to define it as graph convolution. Different from node classification tasks, graph classification tasks need to pay attention to global information of graphs, which requires graph pooling mechanism to extract global information. Recently, Many researchers are devoted to the study of graph pooling and then proposed diversity of graph pooling models. However, in the graph classification tasks, these graph pooling methods are general and the graph classification accuracy still has room to improvement. Therefore, we propose the covariance pooling (CovPooling) to improve the classification accuracy of graph data sets. CovPooling uses node feature correlation to learn hierarchical representation of a graph. Our graph pooling utilizes node information and graph topology. Experiments show that our pooling module can be integrated into multiple graph convolution layers and achieve state-of-the-art performance in some datasets.