Multi-Embedding-based Deep Learning Framework for Fake Review Detection in E-commerce Platforms

The rapid increase in consumer-generated content on e-commerce platforms has heightened the need for reliablemechanisms to verify review authenticity. Traditional approaches to fake review identification have mainly usedsingle-embedding architectures, which are limited in capturing the complex linguistic patterns of fake content. Thisstudy proposes the multi-embedding fake review detector (MEFRD), a detection framework that integrates diversesemantic representations from BERT and RoBERTa transformers to improve feature extraction and classificationaccuracy. The proposed framework was empirically evaluated using a curated dataset of restaurant reviews from theYelp.com, which includes both authentic and fake reviews. MEFRD applies a concatenation-based fusion strategy tocombine textual representations from the two embedding models. A deep neural network classifier then processes theseintegrated features to perform binary authenticity classification. Experimental results show that MEFRD consistentlyoutperforms traditional machine learning and deep learning models. The framework notably demonstrates notableenhancement in F1-score metrics. A comparative analysis of fusion strategies confirms that concatenation-basedintegration outperforms additive and multiplicative methods in preserving semantic richness. Therefore, the findingsvalidate the effectiveness of multi-embedding architectures for fake review detection and provide practical implicationsfor developing trustworthy review verification systems in digital commerce environments.