Feature selection for high dimensionality data in chemical sensing

We present the results obtained by applying different feature selection (FS) methods to the analysis of chemical sensing data. Four datasets have been considered: a mass spectrometer (MS) dataset with 505 peaks (features) and three datasets from metal oxide sensor based E-noses with 30 features. For the MS data we first filter features individually and then apply optimal and suboptimal FS search strategies. The three E-nose datasets present different discrimination hardness and different number of data. For every classification problem three classifiers are tested (3NN, LDA, QDA). FS performance is calculated by two nested cross-validation cycles in order to prevent selection bias. FS always increases the correct test set classification ratio (sometimes the increase is substantial) and the discriminative features are recognized. To our knowledge, this is the first benchmarking study evaluating several FS options (classifiers and search strategies, including suboptimal ones) for the classification of different chemical sensing datasets