High-dimensional clustering using frequency sensitive competitive learning

Abstract

In this paper a clustering algorithm for sparsely sampled high-dimensional feature spaces is proposed. The algorithm performs clustering by employing a distance measure that compensates for differently sized clusters. A sequential version of the algorithm is constructed in the form of a frequency-sensitive competitive learning scheme. Experiments are conducted on an artificial Gaussian data set and on wavelet-based texture feature sets, where classification performance is used as a clustering significance measure. It is shown that the proposed technique improves classification performance dramatically for high-dimensional problems.