Feature Reduction Based on Analysis of Covariance Matrix
Lishi Zhang, Xianchang Wang, Leilei Qu · 2008
This paper presents a novel approach of feature selection based on analysis of covariance matrix of training patterns, a correlation-based feature selection method is put forward. An objective measure is proposed and defined. It is shown that for a given set of features, a subset of features that has the highest sum of the correlation coefficients has the tendency to be reduced, if it meets the requirement of the objective function, a favorable sets is finally retained, when it is omitted, the good classification efficiency is obtained. The algorithm performs elimination, the elimination of which minimizes the value of objective measure, a terminating criterion is given. Experiments show that the proposed algorithm performs well in eliminating irrelevant features while constraining the increase in recognition error rates for unknown data.