Improving recommendation accuracy based on item-specific tag preferences

Abstract

In recent years, different proposals have been made to exploit Social Web tagging information to build more effective recommender systems. The tagging data, for example, were used to identify similar users or were viewed as additional information about the recommendable items. Recent research has indicated that “attaching feelings to tags” is experienced by users as a valuable means to express which features of an item they particularly like or dislike. When following such an approach, users would therefore not only add tags to an item as in usual Web 2.0 applications, but also attach a preference ( affect ) to the tag itself, expressing, for example, whether or not they liked a certain actor in a given movie. In this work, we show how this additional preference data can be exploited by a recommender system to make more accurate predictions. In contrast to previous work, which also relied on so-called tag preferences to enhance the predictive accuracy of recommender systems, we argue that tag preferences should be considered in the context of an item. We therefore propose new schemes to infer and exploit context-specific tag preferences in the recommendation process. An evaluation on two different datasets reveals that our approach is capable of providing more accurate recommendations than previous tag-based recommender algorithms and recent tag-agnostic matrix factorization techniques.