A hybrid fuzzy K-harmonic means clustering algorithm

K-means (KM) clustering is very sensitive to the initialization and easily converges to the local optima. K-harmonic means (KHM) clustering solves this problem by introducing the harmonic averages of the distances as components to its objective function. It is demonstrated through many experiments that KHM is insensitive to the initialization of the cluster centers attributed to a boosting function. However, KHM has a noise sensitivity problem in clustering noisy data because of its probabilistic constraint the same as fuzzy c-means (FCM) clustering. In this paper, we present a hybrid fuzzy K-harmonic means (HFKHM) clustering algorithm based on improved possibilistic c-means clustering (IPCM) and KHM. HFKHM solves the noise sensitivity problem of KHM and improves the memberships of IPCM by combining the merits of KHM and IPCM. The performance of HFKHM is compared with those of KHM and IPCM on several data sets. Experimental results show the superiority of HFKHM.