Feature Shaving for Spectroscopic Data

High-resolution spectroscopy is a powerful industrial tool. The number of features (wavelengths) in these data sets varies from several hundreds up to a thousand. Relevant feature selection/extraction algorithms are necessary to handle data of such a large dimensionality. One of the possible solutions is the SVM shaving technique. It was developed for applications in microarray data, which also have a huge number of features. The fact that the neighboring features (wavelengths) are highly correlated allows one to propose the SVM band-shaving algorithm, which takes into account the prior knowledge on the wavelengths order. The SVM band-shaving has a lower computational demands than the standard SVM shaving and selects features organized into bands. This is preferable due to possible noise reduction and a more clear physical interpretation.