THGFM: Dual-Branch Temporal Heterogeneous Graph Fusion Model

2026-07-29

AI Summary

This work addresses the challenge of jointly modeling structural heterogeneity and temporal dynamics in dynamic heterogeneous graphs by proposing a dual-branch Graph Transformer architecture. One branch employs shared spatiotemporal attention to capture universal dynamic patterns, while the other leverages relation-type-specific temporal attention to model interaction heterogeneity. These branches are adaptively fused via a novel type-conditioned non-competitive gating mechanism that mitigates zero-sum competition between them. The method further introduces rotational temporal attention, which embeds relative time into the query-key matching process through phase rotation. Evaluated across six benchmarks, the approach achieves an average performance gain of 3.25%, with particularly notable improvements of +12.37% on OAG-CS PV, +4.87% on PF-L₂, and +4.24% on OGBN-MAG.

Abstract

Temporal heterogeneous graphs offer a natural abstraction for dynamic relational systems in which diverse node and relation types co-exist and evolve over time. Learning on such graphs requires jointly modeling cross-type structural heterogeneity and the temporal dynamics of interactions, yet existing methods still struggle to reconcile parameter-efficient cross-type transfer with relation-aware specialization, and typically inject time only as additive features outside the attention kernel. We propose \textbf{THGFM}, a web-scale temporal heterogeneous graph fusion model that addresses both limitations within a unified dual-path architecture. THGFM couples a \textit{Shared-Space Temporal Attention} branch for parameter-efficient cross-type transfer with a \textit{Relational Type-Partitioned Temporal Attention} branch for relation-aware specialization, and integrates them through \textit{Dual-Path Relational--Shared Fusion}, instantiated with \textit{Type-Conditioned Non-Competitive Gated Sum Fusion}: a adaptive mechanism that assigns independent, type-conditioned feature-wise gates to the shared and specialized branches, allowing both to be amplified or suppressed without zero-sum competition. To directly incorporate relative time into the attention score, THGFM further introduces \textit{Rotary Temporal Attention}, which rotates queries and keys by half-phases of relative time before matching. THGFM consistently outperforms baseline graph transformer models on academic graphs benchmarks, delivering a $+3.25\%$ six-task mean gain, with peak relative gains of $+12.37\%$ on OAG-CS PV, $+4.87\%$ on PF-$L_2$, and $+1.18\%$ on PF-$L_1$, and $+4.24\%$, $+3.73\%$, and $+4.61\%$ on OGBN-MAG, HTAG-ArXiv, and HTAG-DBLP, respectively.

Problem

• Temporal Heterogeneous Graph
• Structural Heterogeneity
• Temporal Dynamics
• Cross-Type Transfer
• Relation-Aware Specialization

Innovation

• Temporal Heterogeneous Graph
• Dual-Branch Architecture
• Rotary Temporal Attention
• Type-Conditioned Gated Fusion
• Graph Transformer