SOFT SUBSPACE CLUSTERING ENSEMBLE BASED ON HEDONIC GAMES

Clustering ensemble aims at producing consensual, better-quality partitions by integrating base partitions. Hedonic Game based Clustering Ensemble (HCCE) exploits a hedonic game to discover the best coalition structure for all data points. However, the consensus partition obtained by HCCE has much more clusters than the true classes available when the game reaches a Nash equilibrium. Additionally, HCCE evaluates the pairwise similarity by co-association matrix without considering the qualities of base partitions. In this paper, Soft Subspace clustering Ensemble based on Hedonic games (SSEH) is proposed to integrate a set of base partitions generated by a soft subspace clustering algorithm ERKM (Entropy Regularization K-Means). The negative entropy coefficient of ERKM helps to generate diversified base partitions, and a new pairwise similarity is evaluated on the basis of the cluster stability. Moreover, the clusters endowed with Nash equilibrium are merged step by step to the ground truth number of classes by minimizing the loss of the social welfare. The experiment results evaluated by three metrics show that two versions of the proposed SSEH have different advantages in terms of average ranks and W/T/L (Wins/Ties/Losses) on 20 test datasets.