Document clustering around weighted-medoids

In this paper, we propose a new similarity-based k-partitions clustering approach, called CAWP. Given the similarities of pairs of objects in the dataset, CAWP groups these objects into K non-overlaped clusters. Each cluster is represented by multiple objects with different weights, called prototype weight. The more representative an object is with respect to a cluster, the larger prototype weight is assigned to that object in the corresponding cluster. Compared with the traditional k-medoids approach, where each cluster is represented by a single medoid or representative object, the way of using prototype weights to allow multiple objects together to describe a cluster is more appropriate in our view. Experimental study using large document datasets show that CAWP is more favorable than other existing similarity-based clustering approaches as it achieves both good effectiveness and efficiency.