Joint Research Conference
June 24-26, 2014
Robust techniques for high dimensional estimation is a key topic in real-world applications. Most current techniques guard against outliers in the response; however, in many situations outliers can be present also in the covariates. We propose a robust method for regression in the high-dimensional setting that is able to address both kinds of outliers. We develop a custom optimization approach to efficiently fit high dimensional models, and establish connections to classic and Bayesian trimmed regression methods. We also discuss theoretical results characterizing the approach, and present numerical experiments that highlight its relevance.
(Joint work with Aleksandr Aravkin and Aurelie Lozano)