Abstract

Partial Least Squares based Dimension Reduction (PLSDR) is a widely used feature extraction method in bioinformatics and related fields, and model selection is critical to the application of PLSDR. Some previous works fixed an empirical value for the number of extracted dimensions, which is only suitable for some certain data sets. This paper proposes two fast adaptive model selection algorithms, PAS and PIS, for PLSDR based on the regression goodness-of-fit on the training set. Without extra validation, the novel algorithms determine a sub-optimal number of dimensions, which is much better than previous works. Experiments on real microarray data sets demonstrate the high performance of the proposed methods.

Paper Details