On Ambiguity Detection and Postprocessing Schemes using Cluster Ensembles

In this paper, we explore the cluster ensemble problem and propose a novel scheme to identify uncertain/ambiguous regions in the data based on the different clusterings in the ensemble. In addition, we analyse two approaches to deal with the detected uncertainty. The first, simplest method, is to ignore ambiguous patterns prior to the ensemble consensus function, thus preserving the non-ambiguous data as good “prototypes” for any further modelling. The second alternative is to use the ensemble solution obtained by the first method to train a supervised model (support vector machines), which is later applied to reallocate, or “recluster” the ambiguous patterns. A comparative analysis of the different ensemble solutions and the base weak clusterings has been conducted on five data sets: two artificial mixtures of five and seven Gaussian, and three real data sets from the UCI machine learning repository. Experimental results have shown in general a better performance of our proposed schemes compared to the standard ensembles.