Regularized Kernel Local Linear Embedding on Dimensionality Reduction for Non-vectorial Data

Abstract

In this paper, we proposed a new nonlinear dimensionality reduction algorithm called regularized Kernel Local Linear Embedding (rKLLE) for highly structured data. It is built on the original LLE by introducing kernel alignment type of constraint to effectively reduce the solution space and find out the embeddings reflecting the prior knowledge. To enable the non-vectorial data applicability of the algorithm, a kernelized LLE is used to get the reconstruction weights. Our experiments on typical non-vectorial data show that rKLLE greatly improves the results of KLLE.