Researchers Shunli Fu, Jingjing Song, Keyu Liu, Huige Li, and Eric C. C. Tsang have developed HGMamba, a dual-path heterogeneous graph neural network that incorporates the selective state space model Mamba to model complex real-world data with multiple node and edge types.
Heterogeneous graph neural networks frequently face difficulties in balancing the capture of local semantic structures with the modeling of long-range semantic associations, an issue affecting both metapath-based and metapath-free approaches. The study identifies long-range dependency in heterogeneous graphs as a two-level semantic challenge, comprising graph-distance dependency within metapath tokens alongside cross-semantic dependency across metapath tokens.
To address both levels of long-range dependency, HGMamba introduces a metapath-guided selective state space module that serves as the network’s long-range path. The module compresses the multi-hop neighborhood of each metapath into a semantic token and uses Mamba to model joint semantics across tokens in linear time, constructing sequences from metapath-level tokens rather than serialized nodes to preserve type-specific semantic structures.
Operating in parallel, an attention-based local path uses type-specific feature projections and multi-head attention to encode fine-grained neighborhood semantics. An adaptive gating mechanism then dynamically fuses the complementary representations from the local and long-range paths, balancing semantic information across different ranges.
In extensive experiments conducted on four benchmark datasets, HGMamba significantly outperformed state-of-the-art heterogeneous graph neural networks across the vast majority of evaluation metrics and tasks, validating the model’s capacity to integrate multi-scale semantic information.
The motivation for incorporating the selective state space model Mamba centers directly on resolving structural and computational challenges encountered in heterogeneous graph neural network architectures. Conventional approaches categorized under metapath-based or metapath-free paradigms struggle to maintain equilibrium between capturing local semantic structures and tracking long-range semantic associations. By deploying Mamba within a specialized dual-path framework, HGMamba addresses this ongoing trade-off, enabling the modeling of joint semantics across extended sequences in linear computational time.
A key insight underlying the framework is the characterization of long-range dependency in heterogeneous graphs as an intrinsically two-level semantic problem. This structure involves both graph-distance dependency within metapath tokens and cross-semantic dependency across distinct metapath tokens. To capture both dependency tiers without disrupting graph heterogeneity, the metapath-guided selective state space module compresses the multi-hop neighborhood of each metapath into a semantic token. Constructing the sequence from metapath-level tokens rather than serialized nodes ensures that the model systematically preserves type-specific semantic structures while evaluating joint semantics across tokens.
Alongside the long-range state space path, the network operates an attention-based local path designed to capture fine-grained neighborhood semantics. This branch implements type-specific feature projections in combination with multi-head attention, targeting localized structural patterns across different node and edge categories. To unify these divergent perspectives, an adaptive gating mechanism dynamically fuses the complementary representations produced by the local and long-range paths. This dynamic integration balances multi-scale semantic information, preventing local neighborhood details from being overshadowed by global sequence modeling.
The proposed architecture was evaluated through extensive experimental testing across four benchmark datasets representing diverse heterogeneous graph domains. Across these benchmarks, HGMamba demonstrated significant performance advantages over state-of-the-art heterogeneous graph neural networks on the vast majority of evaluation metrics and experimental tasks. The resulting performance gains confirm the efficacy of coupling selective state space sequences with localized multi-head attention, verifying that the dual-path design effectively integrates multi-scale semantic information across complex relational graphs.
Subject of Research: Heterogeneous graph neural networks with selective state space models
Article Title: HGMamba: dual-path heterogeneous graph neural network with selective state space model
Article References: Fu, S., Song, J., Liu, K., Li, H., & Tsang, E. C. C. (2026). HGMamba: dual-path heterogeneous graph neural network with selective state space model. Knowledge and Information Systems, 68(1), Article 307. https://doi.org/10.1007/s10115-026-02930-y
Image Credits: AI Generated
DOI: 10.1007/s10115-026-02930-y
Keywords: heterogeneous graph neural networks, selective state space models, Mamba, graph representation learning, metapath tokens, multi-head attention, adaptive gating, graph neural networks, HGMamba, dual-path, heterogeneous, graph
News Source: Cassandra Pierce. (November 18, 2026). Dual-Path Neural Network Integrates Selective State Spaces for Heterogeneous Graphs. Scienmag.



