Tag Recommendation Based on Continuous Conditional Random Fields

This paper focuses on tag recommendation. In many tagging systems, tags are highly interdependent. Conventional methods do not exploit dependencies between tags when computing the relevance score of a candidate tag for a new resource. In this paper, we take into consideration the relationship between candidate tags and propose a continuous conditional random fields (CRF) model for tag recommendation, referred to as TR-CRF. Firstly, a feature vector is set up for each candidate tag, which contains tag co-occurrence information. Then a ranking model is trained with TR-CRF, and the ranking score represents the relevance between a candidate tag and a resource. With the use of this model, tags of a new resource are generated automatically according to their ranking scores. Experimental results on two real world tag recommendation tasks validate the effectiveness of our method.