Graph convolutional network with multi-similarity attribute matrices fusion for node classification

摘要

Graph convolution networks (GCNs) have become one of the most popular deep neural network-based models in many real-world applications. GCNs can extract features take advantage of both graph structure and node attributes based on convolutional neural networks. Existing GCN models represent nodes by aggregating the graph structure and node attributes from their neighbors which usually disrupt the node similarities in the feature space. In this paper, we propose the MSF-GCN, a graph convolutional network with multi-similarity attributed matrices fusion for node classification. The key idea behind the MSF-GCN is that not only the topology but also the attributes similarities are taken into consideration for node presentation. Specifically, we first apply a GAT-based module to obtain a general representation of the original graph. Next, we construct two k-nearest neighbor graphs based on node attributes with cosine similarity and heat kernel similarity. To balance the disparity between the graph structure and node attributes for each similarity matrix, we develop a self-attention network to integrate the node attributes with topological features. Furthermore, we design a gated-fusion network to merge the cosine similarity vector and heat kernel vector. In our experiments on four real-world datasets, results show that our MSF-GCN model can extract more correlation information from the node attributes and graph structure, and outperform seven state-of-the-art methods.