A New Particle Swarm Optimization Algorithm for Clustering

Cluster analysis is a widely used data mining technique that extracts natural groupings hidden in data to exploit meaningful and comprehensible information. Since the evaluation mechanism based on intra-cluster distance (ICD) function is straightforward, traditional single-objective clustering algorithm fails to handle the data sets with complex distribution for easily resulting in the drop of optimal solutions' accuracy on the late stage of search. To overcome the issue, a novel index reflecting the similarity of data within a cluster is presented and called the intra-cluster cohesion (ICC). From this, we propose a new PSO-based clustering algorithm, whose clustering process comprises two parts. Specifically, the first part is used for minimizing the main objective ICD, and the second part is a fine-tuning process which promotes the clustering accuracy with adopting the criterion of ICC as the new objective. The proposed algorithm has been experimented using six open-source clustering sets with various geometric distributions. The results demonstrate that the new PSO outperforms traditional PSO, KPSO, CPSO and ACPSO in terms of accuracy, and the convergence trends of related algorithms show that the ICC function significantly contributes to the accuracy.