Determining user expertise for improving recommendation performance

Recommender systems are designed to predict user preference for items using his/her past activities. Predominant studies have focused on modeling and developing recommendation algorithm to predict the user preference accurately. In this paper, we assume there are some more reliable and important users for recommendation process, who have deep and broad knowledge of specific domains. Instead of developing a new recommendation model, we propose a method for quantifying user's expertise and exploiting tem to improve performance of existing recommendation algorithms. More specifically, we suggest three general expert factors called early adoption (EA), heavy access (HA) and niche-item access (NA), and we explain how to determine the expertise of each user using a latent variable model. Additionally, we show how our method can be applied to existing recommendation models. On the real-world data from last.fm, our approach shows not only accurate but novel and serendipitous recommendation.