Attention-Based Recurrent Neural Network for Multicriteria Recommendations

Traditional recommender systems use a single rating for each item. However, this approach is limited because a single overall rating does not provide sufficient information about the reasons that led to a user's overall rating. Therefore, multicriteria recommender systems have been developed to benefit from user preferences expressed across variety of criteria. Recurrent neural networks (RNNs) have been also used in recommendation systems; indeed, they have proven their efficiency in other fields as speech recognition and machine translation. Nevertheless, the use of various RNN structures was restricted to reviews and session-based single-criteria recommender systems. These methods require useful metadata, such as the history of user activity during a session. In this paper, we propose a sequence-aware Long-Short Term Memory (LSTM) RNN with a custom attention mechanism to predict the overall ratings of users. The user's multicriteria ratings are the only needed data for the proposed approach. Thus, we consider every user's multicriteria rating given for an item as a sequence of data for that user. Extensive experiments conducted on real-world data show that the proposed method outperforms baseline approaches.