Linear manifold clustering

Abstract

In this paper we describe a new cluster model which is based on the concept of linear manifolds. The method identifies subsets of the data which are embedded in arbitrary oriented lower dimensional linear manifolds. Minimal subsets of points are repeatedly sampled to construct trial linear manifolds of various dimensions. Histograms of the distances of the points to each trial manifold are computed. The sampling corresponding to the histogram having the best separation between a mode near zero and the rest is selected and the data points are partitioned on the basis of the best separation. The repeated sampling then continues recursively on each block of the partitioned data. A broad evaluation of some hundred experiments over real and synthetic data sets demonstrates the general superiority of this algorithm over any of the competing algorithms in terms of stability, accuracy, and computation time. © Springer-Verlag Berlin Heidelberg 2005.