Abstract
In this work, different combinations of dissimilarity coefficients and clustering algorithms are compared in or-der to separate FTIR data in different classes. For this purpose, a dataset of eighty five spectra of four types of sample cells acquired with two different protocols are used (fixed and unfixed). Five dissimilarity coefficients were assessed by using three types of unsupervised classifiers (K-means, K-medoids and Agglomerative Hier-archical Clustering). We introduce in particular a new spectral representation by detecting the signals peaks and their corresponding dynamics and widths. The motivation of this representation is to introduce invariant properties with respect to small spectra shifts or intensity variations. As main results, the dissimilarity mea-sure called Spectral Information Divergence obtained the best classification performance for both treatment protocols when is used over the proposed spectral representation.
