Hybrid-attention mechanism based heterogeneous graph representation learning

Abstract

Heterogeneous graph refers to a type of graph data characterized by its diverse node types and relation types, containing rich structures, features and heterogeneous information. How to fully utilize and capture these key information to generate effective node representations poses a great challenge in heterogeneous graph analysis and mining. To better tackle this problem, a heterogeneous graph representation learning model based on hybrid-attention mechanism is proposed, namely Heterogeneous Graph Relation Attention Network (HGRAN). The main contributions of HGRAN are listed as follows. First, a novel framework was proposed for better representing heterogeneous information originating from various relations and comprehensive usage of both structural and feature information instead of employing meta-path based framework. Second, a novel hybrid-attention mechanism which combines relation attention and node attention was proposed within this framework. Third, a novel feature similarity based relation attention is proposed to capture heterogeneous information originating from different relations. Fourth, in order to better implement node attention in heterogeneous graphs, a new transforming method that transforms adjacency matrices of diverse relations into a unified manner is proposed. Finally, extensive experiments on multiple real-world heterogeneous graph datasets are conducted to verify HGRAN, and the results support its superiority in comparison with the state-of-the-art methods.