A hybrid of XGBoost and aspect-based review mining with attention neural network for user preference prediction

With the rapid development of the internet, users tend to refer to the rating scores or review opinions on social platforms. Most recommendation systems use collaborative filtering (CF) methods to recommend items based on users’ ratings. The rating-based CF methods do not consider users’ review opinions on different aspects of items. The accuracy of the rating predictions can be effectively improved by considering the latent semantics and various aspects of user reviews. In this paper, a novel rating prediction method is proposed according to an attention-based gated recurrent unit (GRU) deep learning model with semantic aspects. A two-phase method is proposed herein; it combines the word attention mechanism and review semantics to extract aspect features from user preferences. In the first phase, a bidirectional GRU neural network is adopted according to word attention in order to extract important words from users’ reviews. In the second phase, we split users’ reviews into words, and generate the aspect-based attention semantic vectors from these reviews based on Latent Dirichlet Allocation and the attention weights of the chosen words. The XGBoost method is then adopted to predict user preference ratings based on the aspect-based attention semantic vectors. The experimental results show that the proposed method outperforms traditional prediction methods and effectively improves the accuracy of predictions.