Novel feature selection method for genetic programming using metabolomic 1H NMR data

Abstract A novel technique for multivariate data analysis using a two-stage genetic programming (GP) routine for feature selection is described. The method is compared with conventional genetic programming for the classification of genetically modified barley. Metabolic fingerprinting by 1 H NMR spectroscopy was used to analyse the differences between transgenic and null-segregant plants. We show that the method has a number of major advantages over standard genetic programming techniques. By selecting a minimal set of characteristic features in the data, the method provides models that are easier to interpret. Moreover the new method achieves better classification results and convergence is reached significantly faster.