Improved least-squares quadratic mutual information clustering via Laplacian Eigenmap

Abstract

Dependence-maximization clustering is another line of clustering framework, which clusters samples by maximizing the statistical dependence on samples in the same group. Recently, dependence-maximization clustering method based on least-squares quadratic mutual information (LSQMI), called LSQMI based clustering (LSQMIC), was proposed. A notable advantage of LSQMIC over other dependence-maximization clustering methods is that it works well even though the data containing outliers. However, the performance of this method tends to decrease in case samples are low density. To deal with this problem, in this paper, we apply Laplacian Eigenmap incorporating with local scaling similarity for representing data so that the samples in the same class will stay as close as possible. Through experiments, we demonstrate that LSQMIC performs better on Laplacian Eigenmap embedded with no losing of the high robustness against outliers.