k-medoids Clustering based on Jensen-Shannon Divergence

Abstract

It is known that k-means and k-medoids clustering are conventional clustering methods. Gen- erally, these methods use the squared L2-norm as a dissimilarity. However, it is difficult to obtain an adequate result by several conventional methods using the squared L2-norm to complex data. It is nec- essary to consider dissimilarity, which considered the structure of the data more precisely. In this paper, we propose JS-divergence based k-medoids, which considers dissimilarity between local distributions. The local distribution is estimated from an object and its neighbor by kernel density estimation(KDE). KDE is a technique to estimate the unknown probability density function, based on a sample of points taken from that distribution. The effectiveness of the proposed method is described by comparing with several conventional clustering methods through numerical experiments. Furthermore, the influence of parameters use in KDE is described. These results showed that the proposed method is effective.