Heterogeneous Network Representation Learning Based on Adaptive Multi-channel Graph Convolution

Abstract

Network representation learning (NRL) is an important technique for network analysis. Heterogeneous information networks (HINs) contain multiple types of nodes and edges, which describe the personalized information of nodes and complex relationships between nodes. In this paper, we propose a Heterogeneous Adaptive Multi-Channel Graph Convolutional Networks (HAM-GCN) for HIN representation learning. In HAM-GCN, we design three channels to extract the specific and common embeddings with respect to each meta-path from node features, topological structures, and their combinations simultaneously. In addition, we design both channel-level attention and semantic-level attention to fuse the low-dimensional representations obtained from different channels and different meta-paths for learning the final representation. A large number of experiments on three benchmark data sets show that HAM-GCN extracts the most relevant information from topological structure and node features, which significantly improves the classification accuracy than other baseline algorithms.