Two-stage semi-supervised clustering ensemble framework based on constraint weight

Semi-supervised clustering ensemble introduces partial supervised information, usually pairwise constraints, to achieve better performance than clustering ensemble. Although it has been successful in many aspects, there are still several limitations that need to be further improved. Firstly, supervised information is only utilized in ensemble generation, but not in the consensus process. Secondly, all clustering solutions participate in getting a final partition without considering redundancy among clustering solutions. Thirdly, each cluster in the same clustering solution is treated equally, which neglects the influence of different clusters to the final clustering result. To address these issues, we propose a two-stage semi-supervised clustering ensemble framework which considers both ensemble member selection and the weighting of clusters. Especially, we define the weight of each pairwise constraint to assist ensemble members selection and the weighting of clusters. In the first stage, a subset of clustering solutions is obtained based on the quality and diversity of clustering solutions in consideration of supervised information. In the second stage, the quality of each cluster is determined by the consistency of unsupervised and supervised information. For the unsupervised information consistency of a cluster, we consider evaluating it by the consistency of a cluster relative to all clustering solutions. For the supervised information consistency of a cluster, it depends on how satisfied a cluster is with the supervised information. In the end, the final partition is achieved by a weighted co-association matrix as consensus function. Experimental results on various datasets show that the proposed framework outperforms most of state-of-the-art clustering algorithms.