ABSTRACT:

Sparsity plays a key role in linear statistical modeling and beyond. In this talk I will discuss the best subset selection problem, a central problem in statistics, wherein the task is to select a set of k relevant features from among p variables, given n samples. I will discuss recent computational techniques relying on integer optimization and first order optimization methods, that enable us to obtain high-quality, near-optimal solutions for best-subsets regression, for sizes well beyond what was considered possible.  This sheds interesting new insights into the statistical behavior of subset selection problems vis-a-vis popular, computationally friendlier methods like L1 regularization -- thereby motivating the design of new statistical estimators with better statistical and computational properties.  If time permits, I will also discuss another closely related, extremely effective, but relatively less understood sparse regularization scheme: the forward stage-wise regression (aka Boosting) in linear models.