Adaptive fusion heterogeneous graph attention network for detecting review spammer groups

Detecting review spammer groups is important because coordinated fraudulent reviewers can jointly manipulate product reputation at a scale that single-review spam detectors often miss. In heterogeneous review graphs, the core challenge is to learn node representations that simultaneously characterize individual-level behavioral abnormality and group-level coordination, and then use them to detect review spammer groups. Existing methods often underemphasize spam-indicative embedding dimensions, combine local behavioral and global structural evidence with a fixed fusion rule, or generate candidate groups from topology alone. We therefore propose AFHGAN, which uses FFAA to recalibrate informative dimensions and adaptively combine local and global evidence, and S-Louvain to jointly consider structural cohesion and embedding similarity during candidate-group construction. On three benchmark datasets, AFHGAN achieves the best P@1000, R@1000, and F1@1000 under the adopted protocol. Relative to original Louvain, S-Louvain improves F1@1000 by 10.8 %, 8.5 %, and 40.0 % on Amazon, YelpChi, and Yelp_four, respectively. These results support the value of retaining behavioral semantics in candidate-group construction.