Explainable fake review detection: A hybrid deep learning model for E-commerce platforms to enhance customer trust

This study addresses the critical challenges in detecting fake reviews on e-commerce platforms by introducing a novel hybrid model that integrates RoBERTa contextual embeddings, TF-IDF features, LSTM networks, and multi-head attention mechanisms. Despite the critical role of online reviews in retail and consumer service decision-making, their credibility is threatened by deliberately induced deceptive content. Studies show that 89% customers check reviews before purchasing, and 47% believe businesses manipulate their online reputation. The proposed model is evaluated using standalone and transfer learning approaches on the OpSpam (hotel reviews) and Deception (hotel, restaurant, doctor reviews) datasets, assessing cross-domain generalization to address these challenges. The results demonstrate the notable performance improvements over existing state-of-the-art methods, with 99.06% accuracy and 99.07% F1-score on OpSpam, 94.01% accuracy and 94.85% F1-score on Deception. Sequential transfer learning substantially improved cross-domain performance, addressing a key limitation in previous approaches. The proposed model uniquely incorporates explainability employing the LSTM hidden-state analysis, Shapley Additive exPlanations (SHAP) value interpretation, and attention-mechanism visualization. In turn, it reveals distinct linguistic patterns between genuine and deceptive reviews across different service domains. This study advances the field of fake review detection by giving solutions that service providers and retailers can use to improve platform integrity and customer trust in a variety of e-commerce scenarios.