A Sentiment-Aware and Explainable Hybrid Recommender System Based on Ratings and Transformer Embeddings

The rapid growth of digital platforms has amplified the demand for intelligent and personalized recommender systems. Traditional collaborative and content-based methods often suffer from cold-start, sparsity, and limited contextual modelling, while transformer-based solutions, though powerful, are computationally demanding. This study proposes a sentiment-enriched hybrid recommender system that integrates Sentence-BERT (SBERT) embeddings, attention-based Aspect-Based Sentiment Analysis (ABSA), and rating-based collaborative filtering methods (SVD, ALS). Two hybridisation strategies were explored: (i) weighted fusion of collaborative and textual signals and (ii) feature concatenation with Gradient Boosting and Shallow Neural Networks. Extensive experiments on IMDb and Amazon datasets show that the hybrid model outperforms both traditional and advanced baselines. The feature-concatenation hybrid achieved the best performance, with an MAE of 0.55, a RMSE of 0.67, and an F1-score of 0.83, representing improvements of up to 25% in MAE and 17% in RMSE over conventional models. Benchmarking against the BERT4Rec transformer baseline confirmed that our approach achieves comparable accuracy with lower computational cost. A controlled user study with 50 participants further validated the system, with the hybrid model receiving the highest satisfaction rating (4.4/5). Additionally, attention visualisation and SHAP analysis enhanced interpretability, making recommendations more transparent. The proposed framework offers a scalable, resource-efficient, and explainable approach to next-generation recommendation, with broad applications in e-commerce, education, and healthcare.