Orthogonal Centroid Locally Linear Embedding for Classification

The locally linear embedding (LLE) algorithm is an unsupervised technique for nonlinear dimensionality reduction which can represent the underlying manifold as well as possible. While in classification, data label information is available and our main purpose changes to represent class separability as well as possible. To the end of classification, we propose a new supervised variant of LLE, called orthogonal centroid locally linear embedding (OCLLE) algorithm in this paper. It uses class membership information to map overlapping high-dimensional data into disjoint clusters in the embedded space. Experiments show that very promising results are yielded by this variant.