Efficient clustering of web-derived data sets

Abstract

Many data sets derived from the web are large, high-dimensional, sparse and have a Zipfian distribution of both classes and features. On such data sets, current scalable clustering methods such as streaming clustering suffer from fragmentation, where large classes are incorrectly divided into many smaller clusters, and computational efficiency drops significantly. We present a new clustering algorithm based on connected components that addresses these issues and so works well on web-type data. © 2009 Springer Berlin Heidelberg.