Efficient Feature Extraction Based on Regularized Uncorrelated Chernoff Discriminant Analysis

In this paper, two regularized uncorrelated Chernoff discriminant analysis (RUCDA) techniques are introduced. As a heteroscedastic extension of the class-wise weighted Fisher criterion, the class-wise weighted Chernoff criterion employed in RUCDA better approximates the Chernoff upper bound of the Bayes classification error in the transformed space, which enable the resulting RUCDA to extract uncorrelated discriminatory information from both mean and covariance differences. Experiments performed on UCI benchmark and protein secondary structure datasets demonstrate good performance of the proposed technique