L0Learn: A Fast Toolkit for Best Subset Selection

L0Learn fits regularization paths for L0-regularized regression. Specifically, it can solve the following class of problems for $q \in \{1,2 \}$: $$ \min_{\beta} \frac{1}{2} || y - X \beta ||^2 + \lambda ||\beta||_0 + \gamma||\beta||_q^q,$$ over a grid of $\lambda$ and $\gamma$ values. Path-wise optimization can be done using either cyclic coordinate descent or local combinatorial search. The core of the toolkit is implemented in C++ and employs many computational tricks and heuristics, leading to very competitive running times compared to popular solvers for the Lasso.

We provide an easy-to-use R interface for L0Learn. For more information on installation and usage, please check L0Learn’s Vignette.

Recent & Upcoming Talks

Fast Algorithms for Best Subset Selection
Dec 16, 2017