Feature selection with prior knowledge improves interpretability of chemometrics models

This paper addresses feature selection for regression of high dimensional data, typically spectrometry measure-ments. In some contexts, prior (but partial) knowledge may be available to guide the selection towards some dimensions a priori assumed to be more relevant. We propose a feature selection method making use of this partial supervision. It extends previous works on feature selection with sparsity-enforcing regularised linear models for classification. In the current regularisation context, a practical approximation of this technique reduces to standard Support Vector Regression learning with iterative re-scaling of the inputs. The scaling factors depend here on the prior knowledge but the final selection may depart from it. Practical results on two data sets show the benefits of the proposed approach on the stability, relevance and interpretability of the selected features, as well as regression performances.