Comparison of Selected Methods for Document Clustering

Abstract

17 cluster analysis techniques proposed for document clustering in terms of internal and external quality measures of clustering and computing time demands are compared. These are combinations of three basic methods (direct, repeated bisection and agglomerative) and five clustering criterion functions for solution assessment (two intra-cluster, one inter-cluster, and two complex ones); all implemented in the CLUTO software package. Furthermore, in the case of the agglomerative method we also applied a single linkage and complete linkage clustering as a criterion function. Collection 20 Newsgroups, a binary vector representation of e-mail messages, was used for comparing the methods. Experiments with document clustering have proved that, from the point of view of entropy and purity, the direct method provides the best results. As regards computing time, the repeated bisection (divisive) method has been the fastest.