Incorporating textual reviews in the learning of latent factors for recommender systems

Abstract

• We use textual reviews to support the ratings in the latent factor model. • A review is interpreted as a description of the user/item and a description of the surrounding elements. • A latent factor model is proposed to account for both interpretations. • Reviews are incorporated not only into objective function but also into the initialization. In the field of recommender systems, the latent factor model is one of the state-of-the-art ones thanks to its strengths in accuracy and scalability. Its core is to learn latent factors for the representation of users and items using rating data collected through surveys after the users have experienced the items. However, on e-commerce applications, besides ratings, users can write reviews for items. A review generally indicates a user’s experience with an item while a rating indicates his/her level of satisfaction with such an experience. Latent factors can be learned more accurately if supported by such reviews. This study is distinctive in interpreting a review as both a description of the user/item and a description of the surrounding elements affecting the user's experience with the item. It has proven to be more effective than those that only consider a review as a description of the user/item. Especially, the analysis of the experimental results shows that our model provides supportive recommendations for users with detailed reviews in spite of their few collected ratings.