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

Area of Science:

• Computer Science
• Data Mining
• Network Analysis

Background:

• Heterogeneous Information Networks (HINs) present complex relationships between diverse node and relation types.
• Existing HIN embedding models often rely on metapaths, limiting their ability to capture high-order relations.
• Metapath structures are pairwise, failing to represent complex interactions like multiple co-authorship.

Purpose of the Study:

• To propose a novel model, Metapath-aware HyperGraph Transformer (Meta-HGT), for learning node embeddings in HINs.
• To address the limitations of existing models in capturing high-order relations within HINs.
• To enhance the accuracy of node classification by effectively utilizing complex relational information.

Main Methods:

• Extending metapaths to guide high-order relation extraction and constructing metapath-based hypergraphs.
• Employing Meta-HGT layers with intra- and inter-hyperedge aggregation, utilizing a type-dependent attention mechanism.
• Fusing node embeddings from multiple hypergraphs using a semantic attention layer for final embedding generation.

Main Results:

• Meta-HGT successfully captures high-order relations in HINs by leveraging hypergraph structures.
• The model learns latent node and hyperedge embeddings through specialized aggregation components.
• Achieved state-of-the-art performance on node classification tasks across three benchmark HIN datasets.

Conclusions:

• Meta-HGT offers a significant advancement in heterogeneous information network embedding.
• The proposed approach effectively models complex, high-order relationships beyond pairwise metapaths.
• Meta-HGT demonstrates superior performance in node classification, highlighting its practical utility.