Jointly Modeling User and Item Reviews by CNN for Multi-domain Recommendation

With the rapid development of e-commerce platforms, the change of shopping ways brings people more choices as well as the overload of the information when shopping. Recommender systems, as the main technology to solve information overload, have been applied to product recommendation successfully. However, the majority of the traditional recommender systems, which are based on the overall rating information of user and focus on single domain recommendation, couldn’t take advantage of user review and the common consumption pattern among multi domains to achieve a better recommendation. In this paper, a multi domain product recommendation algorithm CCoNN based on Convolutional Neural Network (CNN) is proposed, which leverage the common user’s review information among several domains to generate user preference vector and item features vector. After that, such vectors are used to make overall rating prediction for a user-product pair through a Factorization Machine (FM). Experiments show that the algorithm has a better performance than other reviews-based recommendation method.