Embedded locality discriminant GPLVM for dimensionality reduction

Abstract

The Gaussian process latent variable model (GPLVM) had been proved to be good at discovering low-dimension manifold from nonlinear high-dimensional data for small training sets. However, for labeled data, GPLVM cannot achieve a better result because it doesn't use the label information. It turned out to be an effective strategy to employ a discriminative prior over the latent space according to the label information. Existing methods for discriminative GPLVM roughly utilized label information data and ignored the natural structure of the data. In this paper, we embedded the locality discriminative information into the GPLVM which not only preserved the locality of data, but also use the label information of samples. Compared to the discriminative GPLVM, our Embedded Locality Discriminant GPLVM (ELD-GPLVM) introduces a local strategy to extract the discriminative information in local region. Experimental results on UCI datasets show that the proposed algorithm has a good performance on no matter small-scale data sets or a larger-scale dataset.