A robust algorithm for subspace clustering of high-dimensional data

Abstract

Subspace clustering has been studied extensively and widely since traditional algorithms are ineffective in high-dimensional data spaces. Firstly, they were sensitive to noises, which are inevitable in high-dimensional data spaces; secondly, they were too severely dependent on some distance metrics, which cannot act as virtual indicators as in high-dimensional data spaces; thirdly, they often use a global threshold, but different groups of features behave differently in various dimensional subspaces. Accordingly, traditional clustering algorithms are not suitable in high-dimensional spaces. On the analysis of the advantages and disadvantages inherent to the traditional clustering algorithm, we propose a robust algorithm JPA (Joining-Pruning Algorithm). Our algorithm is based on an efficient two-phase architecture. The experiments show that our algorithm achieves a significant gain of runtime and quality in comparison to nowadays subspace clustering algorithms. © 2007 Asian Network for Scientific Information.