A hybrid recommendation framework utilizing domain-adaptive RoBERTa embeddings for enhanced personalization in e-commerce

Abstract

Abstract With the rapid growth of e-commerce and online platforms, delivering personalised and accurate recommendations remains a challenge due to sparse interaction data and diverse user interests. This paper proposes HyReC , a unique hybrid recommendation framework that integrates content-based and collaborative filtering while maintaining computational efficiency. Domain-adaptive RoBERTa embeddings are used to extract semantic representations from textual content, capturing user and item preferences from descriptions and reviews. A Deep Neural Network (DNN) model uses user-item interactions to generate latent behavioural embeddings, which are enriched behavioural statistical features such as mean rating, rating variance (standard deviation), interaction frequency, and skewness. Heterogeneous embeddings are fused using a Bahdanau attention mechanism , enabling the model to dynamically weight content, collaborative, and statistical signals. The fused representation is then used to generate recommendations through a Learning-to-Rank layer , depending on application scale. A model is trained using the Adam optimiser to ensure fast convergence and stable performance. Experimental evaluation on the Amazon Baby dataset demonstrates that HyReC achieves superior performance of 0.15 , MAE of 0.10 , MSE of 0.023 , R² of 0.98 , Pearson Correlation of 0.99 , MAPE of 1.5% , and F1-score of 0.98 , outperforming state-of-the-art models such as LSTM, RBM + KNN, GNN, and GAT. Experiments on benchmark datasets demonstrate that the proposed framework improves recommendation accuracy , diversity , and robustness compared to baseline models, effectively addressing data sparsity , user interest drift , and heterogeneous content .