A Novel Ensemble Clustering Approach with Internal Weighting Strategy

Abstract

Due to its strong robustness and high accuracy, the ensemble clustering has been extensively applied in many fields. Despite the significant achievements, there is a limitation need to be overcome. Most of the existing ensemble clustering methods commonly value all component clusterings equally regardless of their different quality. Therefore, low-quality component clusterings are very destructive for them. Although many modified measures have been adopted to weight each component clustering base on their quality, yet these measures have a tendency to treat each cluster as an unit and ignore the diversity of clusters belonging to same basis clustering. How to take into count the diversity of clusters and the internal structure of component clusterings to improve the consensus performance is a problem that need to be solved. In this paper, a novel ensemble clustering method is proposed to solve this problem. Firstly, We adopt five clustering methods to generate diverse component clusterings. Secondly, in order to build candidate component clustering pool, we propose a new method to discard low-quality component clusterings. Thirdly, we utilize a advanced method to evaluate the unreliability of clusters with respect to candidate component clusterings. Finally, we further utilize a accepted weighting strategy and a consensus function to obtain consensus clustering. We have conducted extensive experiments to demonstrate the superiority of the proposed approach.