The performance factors of clustering ensembles

Abstract

The accomplishments on classifier ensembles originate the studies of clustering ensembles. In this study the factors on performance of clustering ensembles (clustering algorithm, the number of features used in clustering, the size of ensemble, the decision combining algorithm) are investigated and compared on 15 benchmark datasets. The decisions of clustering algorithms based on different feature subsets are combined. On the process of decision combination, the graph partition algorithms are averaged successful while hierarchical algorithms have best individual successes. The number of features used in clustering algorithms increases the success. The size of clustering ensemble is also direct proportional with clustering performance. Kmeans and fuzzy-kmeans are best clustering algorithms over our experimented datasets.