A0891
Title: External information for high-dimensional variable selection
Authors: David Rossell - Universitat Pompeu Fabra (Spain)
Paul Rognon-Vael - U. Pompeu Fabra - U. Politecnica de Catalunya (Spain) [presenting]
Abstract: In many modern applications, the sample size is often insufficient to estimate the parameters of a model reliably. This limitation has motivated the development of high-dimensional techniques that frequently rely on very strict assumptions on sparsity and signal size. Yet, often, external information is available and can be leveraged to enhance parameter estimation. This concept is investigated in the context of variable selection in basic linear regression scenarios. It is shown that incorporating external information allows pushing the theoretical limits under which consistent variable selection is possible. A concrete model selection procedure based on Bayesian principles is proposed, which realizes those benefits and outperforms standard penalization.
