TSL-GFDN: temporally-aware soft structural learning for group-based fraud detection network

The rapid expansion of the digital economy has fostered a complex fraud ecosystem in which group-based attacks, including Sockpuppet-based Targeted Attacks on Reviewing Systems (STARS), manifest as both dense interaction structures and coordinated temporal bursts. Graph neural networks for cohesive subgraph mining have advanced fraud detection, but existing methods often rely on static, discrete core decompositions, such as k- core and ( α , β )-core analyses. These hard structural operations are limited by an inability to model temporal coordination, which ignores the temporal synchronization of fraudulent behavior, and by loss of information near hard thresholds, which limits the model’s use of near-boundary structural signals. We propose T emporally Aware S oft Structural L earning for G roup-based F raud D etection N etwork (TSL-GFDN), a temporally aware extension of GFDN that integrates sliding-window temporal weighting into soft ( α , β )-core feature learning. TSL-GFDN maps temporally weighted degree margins to continuous soft membership scores and uses a temporal risk intensity autoencoder to support community discovery, edge classification, and vertex classification. On the real-world TianChi e-commerce dataset, where fraud exhibits clear spatiotemporal concentration, TSL-GFDN improves F1-score by 13.1% and precision by 25.4% relative to GFDN. On Bitcoin Alpha and Bitcoin OTC, the advantage is more limited because the hub-dominated topology and synthetic attack labels provide little within-density temporal variation. These results show both the value and the boundary of temporal structural learning for group-based fraud detection.