Tag-based smoothing for item recommendation

Collaborative tagging sites allow users to save and annotate their favorite Web contents with tags. These tags provide a novel source of information for collaborative filtering. However, major gap exists in how to integrate tagging information into traditional recommender systems for better recommendation quality, due to the difficulty to quantize these semantic tags. This paper proposes a novel approach to convert this semantic information into quantitative values from a smoothing point of view, taking advantage of the topic-based method, and then make recommendation in a traditional user-based CF fashion based on the smoothed user-item matrix. Experiments on two real-world collaborative tagging datasets prove the effectiveness of our approach.