Meta-HGT: Metapath-aware HyperGraph Transformer for heterogeneous information network embedding

Heterogeneous information network embedding aims to learn low-dimensional node vectors in heterogeneous information networks (HINs), concerning not only structural information but also heterogeneity of diverse node and relation types. Most existing HIN embedding models mainly rely on metapath to define composite relations between node pairs and thus extract substructures from the original HIN. However, due to the pairwise structure of metapath, these models fail to capture the high -order relations (such as Multiple authors co-authoring a paper) implicitly contained in HINs. To tackle the limitation, this paper proposes a Metapath-aware HyperGraph Transformer (Meta-HGT) for node embedding in HINs. Meta-HGT first extends metapath to guide the high-order relation extraction from original HIN and constructs multiple metapath based hypergraphs with diverse composite semantics. Then, Meta-HGT learns the latent node and hyperedge embeddings in each metapath based hypergraph through Meta-HGT layers. Each layer consists of two types of components, i.e., intra-hyperedge aggregation and inter-hyperedge aggregation, in which a novel type-dependent attention mechanism is proposed for node and hyperedge feature aggregation. Finally, it fuses multiple node embeddings learned from different metapath based hypergraphs via a semantic attention layer and generates the final node embeddings. Extensive experiments have been conducted on three HIN benchmarks for node classification. The results demonstrate that Meta-HGT achieves state-of-the-art performance on all three datasets. (c) 2022 Elsevier Ltd. All rights reserved.

Automated summary

This paper proposes a Metapath-aware HyperGraph Transformer (Meta-HGT) for node embedding in heterogeneous information networks (HINs). Meta-HGT extends metapath to guide the extraction of high-order relations and constructs multiple metapath-based hypergraphs with diverse semantics. By using intra-hyperedge and inter-hyperedge aggregation components, along with a novel type-dependent attention mechanism, Meta-HGT learns latent node and hyperedge embeddings in each hypergraph. Finally, multiple node embeddings are fused via a semantic attention layer to generate the final node embeddings.