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 six metal oxide sensors. The 5 features are: the classical relative change in resistance R/R-0; the curve integral both over the gas adsorption and desorbtion process and the phase space integral, again over adsorption and desorbtion. The phase space integral is a novel feature introduced in [1]. We show that performance (in terms of the cross validated test error of a three nearest neighbour classifier) is always significantly better for the best selected 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 univocal regarding the best feature type. Yet, on average over the four datasets in which the complete dataset can be decomposed, the phase integral calculated over the desorption wins. Also, the features (phase and integral) calculated on the desorbtion seem to consistently give higher performance than the corresponding features calculated during adsorption. The standard R/R-0 stands in the lower part of the ranking. (c) 2006 Published by Elsevier B.V.
Comparing the performance of different features in sensor arrays
Pardo M;Sberveglieri G
2007-01-01
Abstract
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 six metal oxide sensors. The 5 features are: the classical relative change in resistance R/R-0; the curve integral both over the gas adsorption and desorbtion process and the phase space integral, again over adsorption and desorbtion. The phase space integral is a novel feature introduced in [1]. We show that performance (in terms of the cross validated test error of a three nearest neighbour classifier) is always significantly better for the best selected 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 univocal regarding the best feature type. Yet, on average over the four datasets in which the complete dataset can be decomposed, the phase integral calculated over the desorption wins. Also, the features (phase and integral) calculated on the desorbtion seem to consistently give higher performance than the corresponding features calculated during adsorption. The standard R/R-0 stands in the lower part of the ranking. (c) 2006 Published by Elsevier B.V.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
