Dynamic Fusion of Topology and History: The AdaGPS Architecture with GatEdgeformer

Abstract

Graph Transformers show immense potential in graph representation learning, yet adaptively capturing complex topological dependencies remains a critical challenge. This paper proposes AdaGPS, a novel adaptive architecture featuring three core improvements. First, the introduced GatEdgeformer module injects edge feature information through a dynamic gating mechanism, enhancing the global attention's perception of complex graph structures. Second, dynamically fusing local structure perception with global attention interaction effectively filters structural noise and accurately extracts long-range dependencies. Third, dynamically fusing historical representations from all previous layers enables the deep AdaGPS architecture to fully utilize rich positional and structural encodings preserved across preceding layers. Extensive experiments strongly demonstrate that the injection of edge features and historical information effectively improves model accuracy. Furthermore, the gated dynamic fusion of local and global features significantly enhances cross-domain generalization capability. Systematic ablations confirm robustness and stability, providing a highly expressive new paradigm for graph representation learning and potential computing power network applications.