A New Deterministic Method of Initializing Spherical K-means for Document Clustering

Abstract

Document clustering is required when the possible categories into which text data are to be organized are not known. Standard clustering algorithms do not suit well due to high sparsity of term matrices of document corpus. Use of cosine similarity among document vector has proved to give good results. Its use with k-means is referred as spherical k-means. The performance of spherical k-means highly depends on its initialization. This paper proposes a deterministic initialization technique for spherical k-means that considers the distribution of vectors within the space. Experiments on real-life data with skewed distributions are done to compare performance with other initialization methods. A related technique to avoid generation of empty clusters is also proposed.