Comparing Cross Validated Classification Accuracies for Alternate Predictor Variable Weighting Algorithms

Abstract

The present research contrasts the effectiveness of four predictor variable weighting algorithms with respect to cross-validated accuracies in classification problems. Ordinary Least Squares Regression (OLS), Ridge Regression (RR), Principle Components (PC), and Logistic Regression (LR), are the techniques that were contrasted on 24 real data sets in terms of optimizing cross-validated classification accuracies. LR was best in only 1 data set, PC was best overall in 16%, RR was best in 8%, and OLS was best in 8% of the data sets.