On evolutionary approaches to unsupervised nearest neighbor regression

Abstract

The detection of structures in high-dimensional data has an important part to play in machine learning. Recently, we proposed a fast iterative strategy for non-linear dimensionality reduction based on the unsupervised formulation of K-nearest neighbor regression. As the unsupervised nearest neighbor (UNN) optimization problem does not allow the computation of derivatives, the employment of direct search methods is reasonable. In this paper we introduce evolutionary optimization approaches for learning UNN embeddings. Two continuous variants are based on the CMA-ES employing regularization with domain restriction, and penalizing extension in latent space. A combinatorial variant is based on embedding the latent variables on a grid, and performing stochastic swaps. We compare the results on artificial dimensionality reduction problems. © 2012 Springer-Verlag.