Selecting Groups of Variables for Prediction Problems in Chemometrics: Recent Regularization Approaches

This presentation addresses the problem of selecting important, potentially overlapping groups of predictor variables in linear models such that the resulting model satisfies a balance between interpretability and prediction performance. This is motivated by data from the field of chemometrics where, due to correlation between predictors from different groups (i.e. variable group “overlap”), identifying groups during model estimation is particularly challenging. In particular, we will highlight some issues of existing methods when they are applied to high dimensional data with overlapping groups of variables. This will be demonstrated through comparison of their optimization criteria and their performance on simulated data. This is joint work in progress with Rebecca Marion, ISBA, Université catholique de Louvain, Belgium.