Double-ConvMF: probabilistic matrix factorization with user and item characteristic text

Abstract

In today’s information-rich society, the importance of recommender systems for matching items and customers is increasing day by day. The development of e-commerce sites and review sites has made it possible to access a large amount of product descriptions and user reviews, and it is believed that more advanced recommendation models can be proposed by efficiently utilizing this text information. ConvMF is the first model that integrates text and probabilistic matrix factorization(PMF) which is one of the matrix factorization methods. In this method, features are extracted from item text such as item descriptions using CNN architecture and integrated into PMF. However they focus only on the item text and not on the user factor. As a result, this method can not reflect user characteristics. Therefore, this paper proposes a new recommender system to extract both item and user features from item and user text using CNN and integrate them into matrix factorization.