An AI-Driven Product Recommendation Framework Integrating Collaborative Filtering and BERT-Based NLP

Recommender systems are essential in enhancing user experience by making precise preference prediction and retrieval of suitable items. Classical collaborative filtering-based approaches typically miss capturing the semantic richness of the textual reviews and break down in ranking quality, generating less customized recommendations. Furthermore, current deep learning-based models like Bert4Rec or composite methods like J-NCFc continue to have high prediction errors and poor ranking accuracy. To fill these voids, this paper introduces a new CF[Formula: see text]BERT model that incorporates collaborative filtering with contextualized review representations obtained from BERT. The originality of this method stems from the integration of user–item interaction patterns and deep semantic representations of reviews to improve both prediction resilience and ranking performance. Experimental outcomes indicate that the model performs excellent rating prediction accuracy with RMSE [Formula: see text] 0.4594 and MAE [Formula: see text] 0.4588 while guaranteeing 92.3% of the predictions within [Formula: see text] tolerance and 100% within [Formula: see text]. When considering ranking tests, the model provides Precision@10 [Formula: see text] 0.423, Recall@10 [Formula: see text] 0.292, and NDCG@10 [Formula: see text] 0.347, further proving its effectiveness in retrieving items that are relevant. Most significantly, the CF[Formula: see text]BERT model achieves a state-of-the-art NDCG [Formula: see text] 0.9534 over baselines including SENT-ROBERTA (0.6403), J-NCFc (0.4065), and Bert4Rec (0.135). These results show that the introduced approach significantly improves recommendation quality, providing both enhanced predictive accuracy and better ranking performance, thus establishing a new benchmark for future recommender system design. The framework has brought forth the possibility of combining context-aware embedding models with collaborative filtering techniques and toward providing intelligent next-generation recommenders. The future extension would be in the direction of multimodal fusion and large-scale deployment to increase adaptability.