A stable Lasso algorithm for inferential sensor structure learning and parameter estimation.

Although the Lasso method has been popular for variable selection in regression modeling, it has been known to yield very different model structures with minor perturbations of the training data. A consequence is that, when cross-validation (CV) is used to determine the hyperparameter λ , seemingly heterogeneous model structures among the CV-folds are resulted for the same λ. In this paper, we propose a new stable Lasso method for model structure learning of static and dynamic models. We begin with building consensus Lasso models with a grid of λ values using all training data. Then the CV-fold models are optimized to conform with the consensus model structures with a modified Lasso objective. In addition, we propose a stable criterion that uses CV errors jointly with a stability measure to select the most stable model with near minimum CV errors. The proposed method is applied to inferential modeling of a chemical plant at DOW Chemical and dynamic modeling of an industrial boiler. • A stable Lasso algorithm is proposed to select variables for inferential sensor modeling. • The algorithm enhances consistent structures in the cross-validation step. • A stable selection criterion of MSE jointly the Jaccard stability measure is proposed. • Successful case studies are presented on inferential modeling of two industrial processes. [ABSTRACT FROM AUTHOR]