Iterative predictor weighting (IPW) PLS: a technique for the elimination of useless predictors in regression problems

Abstract

A new method for the elimination of useless predictors in multivariate regression problems is proposed. The method is based on the cyclic repetition of PLS regression. In each cycle the predictor importance (product of the absolute value of the regression coefficient and the standard deviation of the predictor) is computed, and in the next cycle the predictors are multiplied by their importance. The algorithm converges after 10–20 cycles. A reduced number of relevant predictors is retained in the final model, whose predictive ability is acceptable, frequently better than that of the model built with all the predictors. Results obtained on many real and simulated data are presented, and compared with those obtained from other techniques. Copyright © 1999 John Wiley & Sons, Ltd.