Coarse-to-Fine Robust Heterogeneous Network Representation Learning Without Metapath

Abstract

Influenced by the heterogeneity, representation learning while preserving the structural and semantic information is more challenging for heterogeneous networks (HNs) than for homogeneous networks. Most of the existing heterogeneous representation models are depending on expensive and sensitive external metapaths to help learn structural and semantic information, and they are rarely considering network noise. In this case, a coarse-to-fine robust heterogeneous network representation learning model is proposed without metapath supervision, called CFRHNE. Inspired by the “divide and conquer” idea, the CFRHNE model divides the representation learning process into a coarse embedding stage of learning structural features and a fine embedding stage of learning semantic features. In the coarse embedding stage, a novel type-level homogeneous representation strategy is designed to learn the coarse representation vectors, by converting the heterogeneous structural feature learning of an HN into multiple homogeneous structural feature learning based on node types. In the fine embedding stage, a novel relation-level heterogeneous representation strategy is designed to further learn fine and robust representation vectors, by using the adversarial learning of multiple relations to add the semantic features to coarse representations. Extensive experiments on multiple datasets and tasks demonstrate the effectiveness of our CFRHNE model.