THGFM: New Model for Temporal Heterogeneous Graph Learning

Researchers have introduced THGFM, a Dual-Branch Temporal Heterogeneous Graph Fusion Model, specifically designed to address the complexities of learning on dynamic relational systems. These systems, common in real-world applications, feature diverse node and relation types that change over time, posing challenges for existing graph learning methods in balancing efficiency with specialized understanding. THGFM tackles these limitations through a unified dual-path architecture. It features a "Shared-Space Temporal Attention" branch for efficient cross-type knowledge transfer and a "Relational Type-Partitioned Temporal Attention" branch for specialized, relation-aware processing. These branches are integrated via a "Dual-Path Relational-Shared Fusion" mechanism, which uses type-conditioned non-competitive gated sum fusion to adaptively combine their outputs. Furthermore, THGFM innovates by incorporating "Rotary Temporal Attention," directly embedding relative time into the attention score calculation. The model consistently outperforms baseline graph transformer models across various academic benchmarks, demonstrating significant gains in performance.