OPTIMAL LINEAR AND QUADRATIC CLASSIFIERS FOR TWO-GROUP DISCRIMINANT ANALYSIS

抄録

A simple algorithm was developed for estimating optimal linear and quadratic classifiers (OLC & OQC) for non-normal multivariate predictor variables in two-group discriminant analysis. The algorithm is based on the alternating least squares (ALS) principle. The optimal classifiers compared favorably with the linear and quadratic discriminant function (LDF & QDF) methods in true error rate. Possible generalizations of the optimal classifier approach (ridge regression, robust regression based on the weighted least squares, etc.) were discussed.