Manifold constrained variational mixtures

Abstract

In many data mining applications, the data manifold is of lower dimension than the dimension of the input space. In this paper, it is proposed to take advantage of this additional information in the frame of variational mixtures. The responsibilities computed in the VBE step are constrained according to a discrepancy measure between the Euclidean and the geodesic distance. The methodology is applied to variational Gaussian mixtures as a particular case and outperforms the standard approach, as well as Parzen windows, on both artificial and real data. © Springer-Verlag Berlin Heidelberg 2005.