A social tag recommendation method alleviating cold start based on probabilistic graphical model

Existing social tag recommendation methods suffer from the cold start problem of tags. To this end, a PageRank-Like tags recommendation (MRF-rank) method is proposed. MRF-rank can adjust recommending chances of tags to alleviate cold start. We detect clique-group within communities, which are considered as basic units showing usage patterns of taggings. We present an ensemble Markov random field (eMRF) model to learn the usage patterns of high quality taggings and then estimate qualities of taggings with few usage records. MRF-rank is proposed, which represents estimated qualities as the weights of tag-resource edges. By specifying a preference vector of target user and resource, MRF-rank spreads the weights among vertices, and then generates tags recommendation based on the weights rank of tags. The experimental results comparing existing methods show that MRF-rank achieves better recommendations in terms of recall and precision.