An Attention-Based Recommendation Algorithm

Yan Chu, Shuhao Qi, Yue Yang, Chenqi Shan, Lina Wang, Zhengkui Wang

Abstract: Conventional recommendation algorithms adopt collaborative filtering approach to recommend items to one user through rating history from other users who share similar preferences. However, this method lacks of consideration of the interest levels, as the rating behavior of users varies with different evaluation standards. In this paper, we introduce one Attention-based Recommendation Algorithm (ABRA), which makes full use of users' comment information to make more reasonable recommendations. ABRA facilitates emotion recognition by analyzing the emotions from users' comments and identifying the polarity of emotions by designing an emotional attention-based model. ABRA integrates the emotional attention model into the Recurrent Neural Network (RNN) to train the comments in order to extract the emotion vector, which is further used to classify the comments. Meanwhile, the context attention mechanism is used in the RNN to identify the emotional polarity of each word. This enables model to identify not only the emotion polarity, but also the tendency value of the emotion words for constructing one more accurate and efficient emotion vector. In addition, we adopt the score correcting gate to decide whether the users' rating operation needs to be updated or not. A modified user-item matrix is applied to recommendation and collaborative filtering recommendation is carried out for users according to emotional score. We train the ABRA using movie review data from Cornell University, and test the model using the Amazon Movies&TV and Yelp datasets. The experimental results indicate that, compared with conventional recommendation algorithms, ABRA can more reasonable perceive the users' emotions and make more accurate recommendation.