A Probabilistic Approach to Personalized Tag Recommendation

In this work, we study the task of personalized tag recommendation in social tagging systems. To include candidate tags beyond the existing vocabularies of the query resource and of the query user, we examine recommendation methods that are based on personomy translation, and propose a probabilistic framework for adopting translations from similar users (neighbors). We propose to use distributional divergence to measure the similarity between users in the context of personomy translation, and examine two variations of such divergence (similarity) measures. We evaluate the proposed framework on a benchmark dataset collected from BibSonomy, and compare with two groups of baseline methods: (i) personomy translation methods based solely on the query user; and (ii) collaborative filtering. The experimental results show that our neighbor based translation methods outperform these baseline methods significantly. Moreover, we show that adopting translations from neighbors indeed helps including more relevant tags than that based solely on the query user.