Abstract: Considering the fact that data belong to several clusters to some extent, we import the fuzzy membership of data to clustering analysis, and propose the fuzzy K-Harmonic Means clustering(FKHM) algorithm. The iterative expressions for cluster center and fuzzy membership are deduced respectively. We then describe a unified expression for the iteration of centers, and deduce the conditional probability expression for the centers and data weight functions for FKHM. Finally, the Folkes & Mallows index is used to evaluate the clustering result. Experiment indicates that the fuzzy K-Harmonic Means algorithm can not only overcome the sensitivity to the initial centers, but also improve the quality of clustering results, compared with the K-Harmonic Means.

Key words: fuzzy K-Harmonic Means, cluster center, conditional probability, Folkes & Mallows index