On Lasso refitting strategies (Q2325356)

Defining a refitting strategy as an estimator which reaches a lower mean square error, compared to the first step Lasso solution, it is shown in this paper that the rates for prediction by the refitting procedures are bounded by the Lasso rates plus an 1-norm of difference between Lasso estimator and the refitted vector. Then the authors study the use of previous analyses of the first step Lasso estimation to derive guarantees for refitting strategies. Develops a Boosted Lasso strategy, based on a least-squares refitting with constraints to avoid explosion of the refitted coefficients. Defines also sign consistent refitting as a refitting procedure preserving the signs of the first step Lasso solution and provides Oracle inequalities for sign consistent refitting estimators. Introduces Bregman Lasso, as a generalization of Bregman Iterations, to restrict the possibility to flip signs minimizing MSE mean-while. The results of numerical studies that compare the different approaches in different synthetic and semi-real scenarios complete the paper.