Knowledge and Information Systems Взаимодействие человека и компьютера
HGMamba: dual-path heterogeneous graph neural network with selective state space model
Аннотация
Heterogeneous graph neural networks (HGNNs) play a crucial role in modeling complex real-world data with multiple node and edge types. However, existing methods, including both metapath-based and metapath-free methods, struggle to balance the capture of local semantic structures with the modeling of long-range semantic associations. To address this problem, we propose HGMamba, a dual-path heterogeneous graph neural network equipped with a selective state space model, namely Mamba. HGMamba is motivated by the observation that long-range dependency in heterogeneous graphs is a two-level semantic problem. Specifically, it involves graph-distance dependency within metapath tokens and cross-semantic dependency across metapath tokens. To capture both levels of dependency, we propose a metapath-guided selective state space module. This module compresses the multi-hop neighborhood of each metapath into a semantic token and applies Mamba to model the joint semantics across tokens in linear time, thereby covering both levels of long-range dependency. The resulting sequence is constructed from metapath-level tokens rather than serialized nodes, which preserves the type-specific semantic structure of heterogeneous graphs. In this way, the module serves as the long-range path of HGMamba. In parallel, an attention-based local path employs type-specific feature projections and multi-head attention to encode fine-grained neighborhood semantics. Finally, an adaptive gating mechanism dynamically fuses the complementary representations of the two paths, effectively balancing semantic information across different ranges. Extensive experiments on four benchmark datasets indicate that HGMamba significantly outperforms state-of-the-art HGNNs on the vast majority of evaluation metrics and tasks. These results validate its effectiveness in integrating multi-scale semantic information.