Particle swarm optimisation representations for simultaneous clustering and feature selection

Clustering, the process of grouping unlabelled data, is an important task in data analysis. It is regarded as one of the most difficult tasks due to the large search space that must be explored. Feature selection is commonly used to reduce the size of a search space, and evolutionary computation (EC) is a group of techniques which are known to give good solutions to difficult problems such as clustering or feature selection. However, there has been relatively little work done on simultaneous clustering and feature selection using EC methods. In this paper we compare medoid and centroid representations that allow particle swarm optimisation (PSO) to perform simultaneous clustering and feature selection. We propose several new techniques which improve clustering performance and ensure valid solutions are generated. Experiments are conducted on a variety of real-world and synthetic datasets in order to analyse the effectiveness of the PSO representations across several different criteria. We show that a medoid representation can achieve superior results compared to the widely used centroid representation.