The Permutable k-means for the Bi-Partial Criterion

The bi-partial criterion for clustering problem consists of two parts, where the first one takes into account intra-cluster relations, and the second - inter-cluster ones. In the case of k-means algorithm, such bi-partial criterion combines intra-cluster dispersion with inter-cluster similarity, to be jointly minimized. The first part only of such objective function provides the "standard" quality of clustering based on distances between objects (the well-known classical k-means). To improve the clustering quality based on the bi-partial objective function, we develop the permutable version of k-means algorithm. This paper shows that the permutable k-means appears to be a new type of a clustering procedure.