A Comparative Study of Selective Cluster Ensemble for Document Clustering

Spherical K-means algorithm (SKM) has been widely used for document clustering. Yet it is prone to getting stuck at local optimums. Recently, cluster ensemble has shown to be an effective method in improving the performance of single clustering algorithm such as SKM by combining multiple clustering solutions resulted by SKM. This paper describes and compares three graph partitioning algorithms and four hierarchy clustering algorithms for document clustering, both theoretically and empirically, using a selective cluster ensemble framework. Our experimental results over several document datasets show that, in terms of normalized mutual information, (a) The hierarchy clustering algorithms produce better clustering results than graph partitioning algorithms (b) on most datasets, the best results are arrived when the size of ensemble is between 40 and 60.