Deep Learning–Based Web Mining to Detect Fake Reviews and Improve E-commerce Recommendations

Abstract: This study presents a deep learning–based web mining framework to detect fake reviews and improve e-commerce recommendation quality by integrating textual, behavioral, user–item metadata, and temporal signals. Reviews are modeled as tuples u(i,t,x,s,y) and transformed into structured feature vectors ϕ(r) that concatenate behavior/context features, user and item profiles, rating signals (including deviation from item mean), and time-based features extracted from crawled e-commerce pages and logs. A multimodal fake-review detector combines a neural text encoder (e.g., Transformer) with engineered web-mined features to estimate p(fake∣r) and derive a credibility score c_r=1-p(fake) . This credibility is then used to down-weight suspicious reviews during review aggregation and recommendation learning, enabling a credibility-aware recommender that is more robust to spam and coordinated manipulation. The framework supports joint multi-task optimization of detection and recommendation objectives and evaluates performance using standard detection metrics (Precision/Recall/F1, ROC-AUC, PR-AUC) and ranking metrics (HR@K, NDCG@K).

Keywords: Web mining, fake review detection, deep learning; transformer encoder, credibility scoring, multi-modal fusion, e-commerce recommender systems, joint learning