A Sentiment Enhanced Deep Collaborative Filtering Recommender System

摘要

Recommender systems use advanced analytic and learning techniques to select relevant information from massive data and inform users' smart decision-making on their daily needs. Numerous works exploiting user's sentiments on products to enhance recommendations have been introduced. However, there has been relatively less work exploring higher-order user-item features interactions for sentiment enhanced recommender system. In this paper, a novel Sentiment Enhanced Deep Collaborative Filtering Recommender System (SE-DCF) is developed. The architecture is based on a Neural Attention network component aggregated with the output predictions of a Convolution Neural Network (CNN) recommender. Specifically, the developed neural attention component puts more emphasis on user and item interactions when constructing the latent spaces (user-item) by adding the mutual influence between the two spaces. Additionally, the CNN learns the specific review of users and his sentiments aspects. Hence, it models accurately the item latent factors and creates a profile model for each user. The proposed framework allows users to find suitable items through the comprehensive aggregation of user's preferences, item attributes, and sentiments per user-item pair. Experiments on real-world data prove that the proposed approach significantly outperforms the state-of-the-art methods in terms of recommendation performances.