Particle Swarm Optimization with K-means for Simultaneous Feature Selection and Data Clustering

Clustering is an unsupervised classification task of data mining. The high dimensional data sets generally comprise of irrelevant and redundant features also along with the relevant features, which deteriorate the clustering result. Therefore, to improve the clustering result, feature selection is necessary to select a subset of relevant features to improve discrimination ability of the original set of features. In recent years, evolutionary and swarm intelligence methods have been used extensively to obtain a subset of relevant features by incorporating clustering algorithm; such methods are known as wrapper methods. The Binary Particle Swarm Optimization (BPSO) is a recent and popular swarm intelligence optimization method to perform discrete optimization problem such as feature selection. However, BPSO easily sticks into the local optima due to single information sharing mechanism by the global best particle in the swarm. Therefore, in this paper, we modify BPSO to improve information sharing among particles to avoid local optima by introducing genetic crossover among particles (named as BPSO-X) to produce relevant set of features. We employ BPSO-X and K-means simultaneously to perform feature selection and clustering, respectively, where silhouette index evaluate quality of the selected features. Further, F-measure and Rand index are used as external quality measure for clustering. The performance of the proposed algorithm over competitors is measured in terms of length of the selected subset, relevancy of the selected subset of features, clustering accuracy, robustness, and convergence speed on seven real data sets from the UCI machine learning repository. The BPSO-X outperforms its competitors in terms of all the mentioned criteria.