Heterogeneous ensemble classifier approach for clustering problems

Data clustering plays an important role in many disciplines, including data mining, machine learning and pattern recognition, where learning the inherent grouping structure of data in an unsupervised manner is needed. In this paper, a novel clustering ensemble scheme is presented. It gathers clustering and classification methods in order to increase the clustering performances. The proposed approach firstly evaluate the qualities of all obtained clustering results and selectively chooses part of patterns as good prototype (stable ensemble) and the rest as ambiguous patterns (unstable ensemble). The stable ensemble gathers the patterns belonging to one cluster while the unstable set corresponds to ambiguous patterns located between different clusters. The clustering solution of the stable ensemble is used to train a supervised learner, which is later applied to reallocate the reset of patterns affected to the unstable ensemble. The performance of the proposed scheme is compared to some well-known methods of the literature using several real and simulation examples. The obtained results show that the proposed scheme has a better clustering performance than the compared methods.