Feature evaluation for an electronic nose

We perform feature selection (FS) on an electronic nose (EN) dataset composed of 30 features, obtained by extracting 5 diverse features from the response curves of 6 metal oxide sensors. The 5 features are: the classical relative change in resistance, R/R₀; the curve integral over both the gas adsorption and desorption processes; the phase space integral, again over adsorption and desorption. We show that performance (both classification error and PCA appearance) is always significantly better for the best features than for all 30 features. Moreover - for some of the 5 features types - performance with all 30 features is worse than performance with just the 6 features of a single type. Results are not unequivocal regarding the best feature type. Yet, for 3 out of 4 datasets in which the complete dataset can be decomposed, the phase integral calculated on the desorption wins. Also, features (phase and integral) calculated on the desorption seem consistently to give higher performance than the corresponding features calculated during adsorption. The standard R/R₀ stands in the lower part of the ranking.