PCA-KL: a parametric dimensionality reduction approach for unsupervised metric learning

Abstract

Dimensionality reduction algorithms are powerful mathematical tools for data analysis and visualization. In many pattern recognition applications, a feature extraction step is often required to mitigate the curse of the dimensionality, a collection of negative effects caused by an arbitrary increase in the number of features in classification tasks. Principal Component Analysis (PCA) is a classical statistical method that creates new features based on linear combinations of the original ones through the eigenvectors of the covariance matrix. In this paper, we propose PCA-KL, a parametric dimensionality reduction algorithm for unsupervised metric learning, based on the computation of the entropic covariance matrix, a surrogate for the covariance matrix of the data obtained in terms of the relative entropy between local Gaussian distributions instead of the usual Euclidean distance between the data points. Numerical experiments with several real datasets show that the proposed method is capable of producing better defined clusters and also higher classification accuracy in comparison to regular PCA and several manifold learning algorithms, making PCA-KL a promising alternative for unsupervised metric learning.