Selection of optimum training sets for use in pattern recognition analysis of chemical data

Abstract

An algorithm is described for the automated selection of optimum training sets for use in pattern recognition studies. A direct calculation is implemented that ensures the data space is sampled equitably in the construction of the training sets. Through this approach, the variety of data in the training set can be maximized while keeping the number of patterns to a specified minimum. A large volume of passive Fourier transform infrared (FT-IR) remote sensing data is used to show the utility of the technique. The algorithm is used to select numerous training sets of various sizes. These training sets are used to develop linear discriminants for pattern recognition. The performance of the discriminants developed from each training set is subsequently evaluated based on their ability to classify a large set of patterns not included in the training procedure. The results are also compared with the performance of discriminants developed using numerous randomly selected training sets. The training sets selected with the new algorithm produce pattern recognition results which are markedly superior to those produced by the randomly selected training sets. © 1991.