Multimodal machine learning framework for fake review detection

Abstract

Online reviews significantly influence consumer decision-making, yet their credibility is increasingly undermined by the rise of fake and manipulated content. This study addresses the growing challenge of detecting deceptive online reviews by developing a highly accurate, robust, and explainable machine learning framework that supports trust and reliability in digital marketplaces. The proposed multimodal framework integrates textual, behavioural, temporal, and network-based features to enhance detection performance. Textual characteristics are extracted using term frequency-inverse document frequency (TF-IDF) and sentiment analysis, while behavioural and temporal attributes model reviewer activity patterns. Network-oriented features capture suspicious reviewer interactions. To mitigate class imbalance, synthetic samples are generated using the synthetic minority over-sampling technique (SMOTE). Several machine learning models—including logistic regression, decision trees, XGBoost, and a stacking ensemble—are trained and evaluated. Experimental findings show that XGBoost and the stacking ensemble deliver strong balanced performance, achieving an F1-score of approximately 0.87 and an accuracy of 0.94. Decision Trees exhibit high precision (0.98), albeit with comparatively lower recall. To ensure transparency and interpretability, Shapley additive explanations (SHAP) are used to analyse model predictions. Results indicate that reviewer connectivity, co-reviewer counts, and sentiment–rating inconsistencies are among the most influential features. Overall, the proposed framework enhances detection accuracy and provides meaningful, explainable insights, making it well-suited for deployment in real-world digital marketplaces. Future work will focus on extending the framework to multilingual datasets and incorporating adaptive learning mechanisms to address evolving deceptive behaviour.