Mixed Multi-channel Graph Convolution Network on Complex Relation Graph

Abstract

Graph Convolutional Networks (GCNs) have attracted increasing attention on network representation learning because they can generate node embeddings by aggregating and transforming information within node neighborhoods. However, the classification effect of GCNs is far from optimal on complex relationship graphs. The main reason is that classification rules on relationship graphs rely on node features rather than structure. most of the existing GCNs do not preserve the feature similarity of the node pair, which may lead to the loss of critical information. To address these issues, this paper proposes a node classification network based on semi-supervised learning: Mif-GCN. Specifically, we propose a mixed graph convolution module to adaptively integrate the adjacency matrices of the initial graph and the feature graph in order to explore their hidden information. In addition, we use an attention mechanism to adaptively extract embeddings from the initial graph, feature graph, and mixed graph. The ultimate goal of the model is to extract relevant information for improving classification accuracy. We validate the effectiveness of Mif-GCN on multiple datasets, including paper citation networks and relational networks. The experimental results outperform the existing methods, which further explore the classification rules.