Improvement of Recommender Systems based on Reviews using Neural Attention Mechanism and LSTM

Recommender systems on online sales sites usually collect users' opinions about the products in two ways namely rating and reviews. In the reviews content, there is a lot of information that is less commonly used and they also differ in importance. This paper presents a review-basedrecommender system using a deep learning approach and the attention mechanism. This model consists of two parallelnetworks, one is trained to model the users and the other one to model the items. Each of these networks comprises the fourphases of: preprocessing, word embedding, feature extraction, and the attention mechanism. Then, in the last layer, the twonetworks are merged and with matrix factorization method, the final estimated rating is obtained. Simulation results of theproposed model are compared with two other models, namely DeepCoNN and NaRRe, and show that the proposed modelperforms better in terms of RMSE and MAE evaluation metrics.

Keywords:

text processing , recommender systems , LSTM , attention mechanism