Regularization Methods for High-Dimensional Data

Estimation with ℓ1-penalty, also known as the Lasso, has become extremely popular, as the method is tailored for high-dimensional data, and is computationally very attractive. The theoretical properties of the standard Lasso are by now quite well understood. We will develop new theory for nonlinear models, such as in generalized-linear and semi-parametetric models. One of the goals will be to gain more insight into the variable selection properties of the Lasso and its modifications, such as the group Lasso and the adaptive Lasso. We moreover will further develop the restricted eigenvalue conditions used to prove oracle results, and to refine the arguments for the case of highly correlated design.

DFG-Verfahren Forschungsgruppen

Teilprojekt zu FOR 916:  Swiss-German Bilateral Research Unit on: Statistical Regularisation and Qualitative Constraints - Inference, Algorithms, Asymptotics and Applications

Internationaler Bezug Schweiz