Measuring and addressing the impact of cold start on associative tag recommenders

Tag recommendation methods that exploit co-occurrence patterns of tags have consistently produced state of the art results. However, tags are not present in significant portions of Web 2.0 objects, which may impact the effectiveness of such methods. This problem, known as cold start, is the focus of this paper. We here evaluate the impact of the cold start on a family of methods for recommending tags. Our results show that the effectiveness of these methods suffer greatly when they cannot rely on previously assigned tags in the target object and that the use of automatic filtering strategies to alleviate the problem yields limited gains. We then propose a new strategy that exploits both positive and negative relevance feedback (RF) from the users to iteratively select input tags to these methods. The results show that the proposed strategy generates significant gains (up to 45%) over the best considered baseline. It is also shown that the proposed method is robust to the lack of user cooperation.