Feature Reduction Using Locally Linear Embedding and Distance Metric Learning

Abstract

Feature reduction is an important issue in pattern recognition. Lower feature dimensionality reduces the classifier complexity and enhances the generalization ability of classifiers. In this paper, researchers proposed a new method for feature dimensionality reduction based on Locally Linear Embedding (LLE) and Distance Metric Learning (DML). Researchers first adopt metric learning method to enhance the class separability, and map the original data to a new space. They use a transformation learned from the data via metric learning method, and then utilize the LLE method to generate an embedding from the transformed data to a lower dimensional manifold. Thus they achieving feature dimensionality reduction, where the final mapping for feature reduction is the composition of the above two transformations learned via DML and LLE method respectively. The method introduces the LLE method traditionally used in unsupervised tasks into the supervised learning domain via a proper and natural way. Experiment results clearly demonstrate the efficiency of the proposed feature reduction method in supervised learning tasks.