Combined Metapath Based Attention Network for Heterogenous Networks Node Classification

Abstract

In recent years, Graph Neural Networks(GNNs) have been widely used as representation learning methods on graphs especially homogeneous graphs, and demonstrated remarkable performance in various tasks. However, GNNs on Heterogeneous Graphs(HGs) haven’t been fully explored, and existing methods on HGs use either metapaths to extract semantics on generated graphs or construct attention mechanics to deal with the original graph directly. The former methods strongly depend on metapaths which makes their performance unstable, and the latter ones can hardly capture deep patterns on HGs as metapaths do. In this paper, we classify information between HG nodes into two parts, prior node information and direct node information, and propose a Combined metapath based Attention Network(CAN) to combine them that making up each one’s disadvantages. Moreover, any number of metapaths can be used in CAN which makes the proposed method more flexible. Based on metapaths we extract the prior node information, and with a novel attention mechanism, we extract the direct node information. Through additional semantic-level attention, we combine them into unique representations. Node classification experiments on real-world datasets demonstrate the performance of the proposed method.