A Recommender System using Tag-based Collaborative User Model

Abstract

Internet users are overwhelmed by a huge media amount available online. Therefore there is a need of an automated way to make compelling recommendations to users according to their needs. There have been many research efforts to reduce that huge amount of content to what the user really needs or prefers. Recommender systems assisting users in easily finding the useful information, are a main research topic that serves this area. According to techniques recommender systems employ, they are mainly classified into three categories: a collaborative-based filtering, content-based filtering, and hybrid filtering. Collaborative filtering relies on the collaboration of users by capturing their judgments on items, and then recommends these items to users with similar taste. Content-based filtering takes advantage of content of a user's preferred items and recommends new items that have similar content. Hybrid filtering takes advantage of both collaborative and content- based filtering and might be in a different ways. No matter what the technique is used, recommender systems require an accurate user model that can reflect a user's characteristics, preferences, and topics of interest. In addition, the systems should take into account users who newly join the systems and thus has presented few opinions, commonly referred to as the cold start users problem. In our research, by leveraging user-generated tags, we propose the topic-driven enriched user model (EM), which is a new way of modeling a user's topics of interest in collaboration with other similar users, in order to improve the recommendation quality and alleviate the cold start user problem. We also present how the proposed model is applied to item recommendations by using locally weighted naive Bayes approach. For evaluating the performance of our model, we compare experimental results with a user model based on user-based collaborative filtering, a user model based on an item-based collaborative filtering, and a vector space model. The experimental results shows that EM outperforms the three algorithms in both recommendation quality and the cold start situation.