Indexed on: 09 Feb '16Published on: 09 Feb '16Published in: Natural computing
A population-based clustering technique, which attempts to integrate different particle swarm optimizers (PSOs) with the famous k-means algorithm, is proposed. More specifically, six existing extensively studied PSOs, which have shown promising performance for continuous optimization, are hybridized separately with Lloyd’s k-means algorithm, leading to six PSO-based clustering methods. These PSO-based approaches use different social communications among neighbors to make some particles escape from local optima to enhance exploration, while k-means is utilized to refine the partitioning results for accelerating convergence. Comparative experiments on 12 synthetic and real-life datasets show that the proposed population-based clustering technique can obtain better and more stable solutions than five individual-based counterparts in most cases. Further, the effects of four different population topologies, three kinds of parameter settings, and two types of initialization methods on the clustering performance are empirically investigated. Moreover, seven boundary handling strategies for PSOs are firstly summarized. Finally, some unexpected conclusions are drawn from the experiments.