Integrating transformer-based credibility signals into neural collaborative filtering for fake review-aware recommendation

Abstract

Abstract Online recommender systems (RS) face growing trust challenges as deceptive reviews distort user feedback. Although RS optimisation and fake review detection have advanced separately, integrating credibility signals directly into recommendation training remains underexplored. This study proposes the Fake-Review-Aware Recommender System (FRARS), which embeds transformer-based deception probabilities into the training objective of a Neural Matrix Factorisation (NeuMF) model. Among several detectors, DeBERTa-v3-base performed best (ROC-AUC = 0.932 on YelpCHI, 0.921 on YelpNYC). FRARS applies these probabilities through two mechanisms: Hard Filtering removes interactions above a deception threshold, while Soft Weighting proportionally down-weights uncertain ones. We evaluate FRARS on two independent Yelp datasets—YelpCHI (67,395 reviews, ~ 49% deceptive) and YelpNYC (359,052 reviews, ~ 10% deceptive). FRARS-Soft improved NDCG@10 by 20.9% on YelpCHI and 19.5% on YelpNYC, with parallel gains in precision and recall; all transformer-based improvements were statistically significant ( p  < 0.001). Detector quality and recommendation gains exhibited a significant monotonic relationship (Spearman ρ = 0.964 on YelpCHI, 0.929 on YelpNYC). These consistent results across different regions, scales, and deception levels indicate that FRARS offers a practical, modular pathway toward more trustworthy recommendation platforms.