A KPSO-based divisive hierarchical clustering algorithm

Classical unsupervised clustering algorithms (e.g. K-Means, EM) are sensitive to initial starting conditions and are difficult to predefine the number of clusters (K). To tackle these problems, a hybrid PSO (Particle Swarm Optimization) and K-Means algorithm, named KPSO clustering, is presented with high performance and without the sensitiveness to the initial starting centers. And then, a KPSO-based Divisive luerarchical Clustering (KPSODHC) method is presented, which aims at an adaptive proper K by terminating the operations of splitting clusters into small ones according to the MinMaxCut threshold. The proposed algorithms are applied to real data-set in contrast to classical K-Means clustering, and the experimental results show that both are effective, and the latter has much potential for its robustness and reliability.