RECOGNITION OF CHEMICAL MARKERS IN CHROMATOGRAPHIC DATA BY AN INDIVIDUAL FEATURE RELIABILITY APPROACH TO CLASSIFICATION

A discriminant analysis algorithm based on the classification reliability of individual features was applied to GC and GC-MS data. Two feature selection methods were investigated in which feature weights were defined either by Fisher ratios or by classification accuracy in single feature dimensions. The discriminant function employed was a linear function of these feature weights where the coefficients of summation were reliability values, ranging from -1 to +1, measuring the confidence with which classification can be made on the basis of that feature alone. The two feature selection methods used in combination with this classifier were competitive with the classical discriminant approach of Fisher on the data sets tested. This approach provides information useful for the recognition of unique chemical markers characteristic of particular groups of samples. The emphasis on chemical markers is important because wider use of pattern recognition with GC and GC-MS requires understanding the chemical basis for discrimination.