Principal Component Analysis Based on Nonparametric Maximum Entropy 1.0.0

Ran He — Fri, 02 Dec 2011 05:45:02 -0000

In this paper, we propose an improved principal component analysis based on maximum entropy (MaxEnt) preservation, called MaxEnt-PCA, which is derived from a Parzen window estimation of Renyi’s quadratic entropy. Instead of minimizing the reconstruction error either based on L2-norm or L1-norm, the MaxEnt-PCA attempts to preserve as much as possible the uncertainty information of the data measured by entropy. The optimal solution of MaxEnt-PCA consists of the eigenvectors of a Laplacian probability matrix corresponding to the MaxEnt distribution. MaxEnt-PCA (1) is rotation invariant, (2) is free from any distribution assumption, and (3) is robust to outliers. Extensive experiments on real-world datasets demonstrate the effectiveness of the proposed linear method as compared to other related robust PCA methods.

mloss.org Principal Component Analysis Based on Nonparametric Maximum Entropy

Principal Component Analysis Based on Nonparametric Maximum Entropy 1.0.0