Incremental Tag-Aware User Profile Building to Augment Item Recommendations

Folksonomic system allows users to use tags to describe items. These tags do not just exist in the form of textual description, and they actually bear more meaning underneath, such as user preference. In this paper, we first show the distribution of preferences and semantic categories across a folksonomic system, and then develop a hybrid design to cope with the cold-start problem. Specifically, we speculate that the semantic categories formed in users‟ perspective and in items‟ perspective are different. They represent different preferences and meaning and are believed to be crucial in recommender algorithm design. Through a dimensionality reduction technique, the Latent Dirichlet Allocation, we demonstrate our speculation is correct. In this regards, we design a hybrid strategy for a movie recommender system. Our system consists of two core modules, namely tag recommendation and profilebased item recommendation. The former module leverages user-tags and item-tags to generate recommended tags, whereas the later one enables the use of the mixture of tags and keywords during the recommendation process. It is worth noting that such design requires no prior information of users and hence the cold-start problem prevented. Author