SBERT-RF: A Scalable Model for Fake Review Detection

Fake reviews have become a persistent problem on restaurant and hospitality platforms, quietly eroding the trust consumers place in peer ratings. Most existing detection approaches rely solely on either text-based natural language processing or simple behavioral heuristics, leaving meaningful predictive signal untapped. This paper introduces ReviewScan, a three-stream ensemble that processes each review through three parallel channels: (i) 384-dimensional dense embeddings from a Sentence-BERT (SBERT) transformer; (ii) 27 behavioral features drawn from reviewer profile metadata, including 10 novel interaction terms developed specifically for this work; and (iii) 32 scalar outputs from a Mamdani fuzzy inference engine that converts expert suspicion rules into continuous scores. All three channels are concatenated into a single 443-dimensional vector and fed to a class-balanced Random Forest of 300 trees. On the publicly available Yelp Chicago restaurant review benchmark of 26,955 labeled samples, ReviewScan achieves 88.63% accuracy and an AUC-ROC of 0.9393, outperforming TF-IDF baselines, a standalone SBERT model, and behavioral-only classifiers. A seven-way ablation study confirms statistically significant, non-overlapping contributions from each channel (p < 0.05, McNemar’s test). The full pipeline runs as a REST API and completes end-to-end inference in 152 ms on standard CPU hardware.