MulAttenRec: A Multi-level Attention-Based Model for Recommendation

It is common nowadays for online buyers to rate shopping items and write review text. This review text information has been proven to be very useful in understanding user preferences and item properties, and thus enhances the capability of Recommender Systems (RS). However, the usefulness of reviews and the significance of words in each review are varied. In this paper, we introduce a multi-level attention mechanism to explore the usefulness of reviews and the significance of words and propose a Multi-level Attention-based Model (MulAttRec) for the recommendation. In addition, we introduce a hybrid prediction layer that model the non-linear interaction between users and items by coupling Factorization Machine (FM) to Deep Neural Network (DNN), which emphasizes both low-order and high-order feature interaction. Extensive experiments show that our approach is able to provide more accurate recommendations than the state-of-the-art recommendation approaches including PMF, NMF, LDA, DeepCoNN, and NARRE. Furthermore, the visualization and analysis of keyword and useful reviews validate the reasonability of our multi-level attention mechanism.