Development of a Supervised non-negative Matrix Factorization (NMF) method for multivariate data with application to chemical kinetics data

Monitoring chemical reactions is crucial in the pharmaceutical and chemical industry. To do so, one can use on-line techniques such as infrared spectroscopy. This technology produces high dimensional data in the form of spectra recorded over the duration of the reaction. The NMF (nonnegative matrix factorization) is particularly effective in decomposing the spectra in an interpretable way. On the other hand, this method is unsupervised and does not allow to derive a prediction model. Thus, in this master's thesis a supervised penalty term is added into the NMF objective function and the update rules are derived. Then, this new algorithm is applied on real and artificial datasets, along with existing unsupervised and supervised methods (PCA, NMF, PLS) which already proved their efficiency for this class of data. The interpretability and the quality of prediction are compared for those methods.