1 Presentation

• 
  02:00 PM - 03:30 PM

  Outlier-Robust Estimation for High Dimensional Statistics

  • Constantine Caraminis, presenter, University of Texas

  This tutorial is about a classical problem in a new regime: dealing with missing, noisy and corrupted data, in the setting of high-dimensional data, where we may have more parameters to estimate than samples available. This high-dimensional regime has been largely driven by application areas such as behavior prediction, collaborative filtering, e-commerce, bioinformatics and genomics, to name just a few. Common to all these is that available data may often be corrupted, subject to noise, erased and missing, or possibly even maliciously (manipulatively) corrupted.

  The goal of this tutorial is twofold. First, we aim to convey the statistical and also algorithmic challenges of robust estimation in high dimensional data. Next, focusing on the two fundamental problems of regression and covariance estimation, we describe several new advances, outlining the development of convex optimization-based algorithms, as well as more efficient greedy-based algorithms, thus essentially describing the state of the art in high-dimensional robust statistics for this class of problems.