TAHAN: Type-Aware Heterogeneous Graph AttentionNeural Networks

Authors:

Abstract
Learning from heterogeneous graphs remains challenging when structural complexity surpasses the modeling capacity of existing heterogeneous graph neural networks (HGNNs). Although current HGNNs capture structural and semantic relations, they typically treat node types as static categorical labels and overlook the contextual influence these types exert during message passing. Metapath-based models further restrict aggregation to the endpoints of a metapath, neglecting the semantic contribution of intermediate node types. As a result, they underexploit type-dependent contextual signals and often struggle in settings with sparse or weakly connected entity types. We propose TAHAN (Type-Aware Heterogeneous Graph Attention Network), a model that explicitly incorporates type-sensitive and context-dependent information into heterogeneous neighborhood aggregation. TAHAN first gathers neighbors across full metapath instances and multiple neighborhood levels, organizing them into type-specific neighborhoods. It then applies a two-stage attention mechanism: type-specific attention to capture contributions from nodes within the same type, followed by type-aware attention to model interactions across different types. The resulting representations fuse these signals into expressive, type-sensitive node embeddings. Experiments on multiple benchmarks show that TAHAN consistently outperforms state-of-the-art HGNNs. Component-level analyses further demonstrate the interpretability and practical value of type-aware modeling, offering insights for future research.

This publication has 15 references indexed in Scilit: