Clustering ensemble learning method based on incremental genetic algorithms

Over the past decade, the clustering ensemble has been emerged as a prominent method as far as the improving of clustering accuracy is concerned. Two major difficulties in clustering ensemble include diversity of clustering and consensus functions. Genetic algorithms are well known methods with high ability to resolve optimization problems including clustering. So far, limited genetic-based clustering ensemble algorithms have been developed. However, their clustering accuracy and convergence to group unlabeled samples are not still satisfied. Generally, associated common problems in traditional genetic algorithms include lose population diversity, clustering invalidity, and context insensitivity. In order to address the above mentioned challenges, this study is devoted towards the development of a clusterer and a clustering ensemble learning method based on incremental genetic algorithms addressing group unlabeled samples. Firstly, an architecture for the clustering ensemble based on incremental genetic-based algorithms is proposed consisting of two phases: (i) to produce cluster partitions as initial populations, (ii) to combine cluster partitions and to generate final clustering solution by incremental genetic based clustering ensemble learning algorithm. In the first and second phases, a threshold fuzzy c-means clustering algorithm as a clusterer and a pattern ensemble learning method based on the incremental genetic-based algorithms are proposed respectively. In the first phase, the quality of cluster partitions belonging to initial populations is measured, in terms of diversity and clustering accuracy. In the second phase, the performance of incremental genetic-based clustering ensemble algorithms is measured, in terms of clustering accuracy and convergence. A comprehensive experimental analysis is conducted by several experiments to evaluate the performance of the proposed clusterer and incremental genetic-based clustering ensemble algorithm which has been tested on the twelve benchmark datasets. In comparison to different clusterers, experimental results show that the proposed clusterer is able to produce cluster partitions with various diversity and desirable clustering accuracy. Moreover, experiments demonstrate that final clustering solution generated by the proposed incremental genetic-based clustering ensemble algorithm using the pattern ensemble learning method possess comparative or better clustering accuracy than clustering solutions generated by the incremental genetic-based clustering ensemble algorithms using other recombination operators. In addition, experiments prove that incremental genetic-based clustering ensemble algorithm speed up to converge into an optimal clustering solution, where pattern ensemble learning method and the cluster partitions produced by the threshold fuzzy c-means clustering algorithm are employed as recombination operator and initial population, respectively.

Preview PDF FSKTM 2012 8R.pdf Download (878kB) | Preview

Actions (login required)

View Item