A0839
Title: Sparse high-dimensional regression with discrete optimization
Authors: Peter Radchenko - University of Sydney (Australia) [presenting]
Abstract: Recent applications of discrete optimization techniques are discussed in high-dimensional regression, concentrating on the algorithmic framework for grouped variable selection. While there exist appealing approaches based on convex relaxations and nonconvex heuristics, we will focus on optimal solutions for the L0-regularized formulation, a problem that is less explored due to computational challenges. The proposed methodology covers nonparametric sparse additive modelling with smooth components and allows for pairwise interactions. Experiments based on the US Census Planning Database demonstrate that our methods automatically identify useful interactions among key factors that have been reported in earlier work by the US Census Bureau. In addition to being useful from an interpretability standpoint, our models lead to predictions that are comparable to popular black-box machine learning methods based on gradient boosting and neural networks.