- Wiki can be found here
- Bug-tracking can be found here
- Install conveniently via pip install git+https://github.com/nicococo/AdaScreen.git
We developed our screening framework in the Python programming language and it can be freely downloaded using the link above or conveniently downloaded and installed automatically, using the pip command.
It is designed to efficiently implement various screening rules without changing the lasso path solver (e.g., scikit-learn lasso solver). Even though different screening rules require different constraints and equations, they all share common data structures; thus, we wrap all of them into a single framework. An advantage of this approach is that the lasso path solver needs to interact with only one abstract class for screening rules.
To systematically manage data structures involved in screening, we divide them into Globals and Locals, where Globals refer to variables that do not change over the lambda path (e.g., the inputs X, y, lambda_max). In contrast, Locals refer to variables that change over the lambda path (e.g., the last regularization parameter lambda_0 or the solution beta^*(lambda_0) from the previous lambda_0). We maintain these data structures in the path solver, and call the init method in the ScreeningRule before entering the main iteration loop of our screening framework.
Furthermore, we designed our screening framework in such a way that all screening rules can be derived from the abstract base class. For example, many screening rules can be framed with a single sphere constraint, consisting of a center and a radius. In such cases, screening rules can be implemented by overloading the getSphere function. For more advanced methods, corresponding functions need to be overloaded. Now, let us consider how AdaScreen can be instantiated under our framework. AdaScreen maintains a list of ScreeningRules itself and can return global and local half-space constraints. Therefore, it is easy to implement AdaScreen with any sphere and any half-space constraints. For example, to implement AdaScreen with EDPP sphere constraint and Sasvi local half-space constraint, we first instantiate EDPP, Sasvi, and AdaScreen. Then in AdaScreen, we simply call setSphereRule(EDPP) and addHalfspace(Sasvi).
In order to solve large-scale lasso problems, screening algorithms have been developed that discard features with zero coefficients based on a computationally efficient screening rule. Most existing screening rules were developed from a spherical constraint and half-space constraints on the dual optimum. However, existing rules admit at most two half-space constraints due to the computational cost incurred by the half-spaces, even though additional constraints may be useful to discard more features. In this paper, we present AdaScreen, an adaptive lasso screening rule ensemble, which allows to combine any one sphere with multiple half-space constraints on a dual optimal solution. Thanks to geometrical considerations that lead to a simple closed form solution for AdaScreen, we can incorporate a large number of half-space constraints at small computational cost. In our experiments, we show that AdaScreen with multiple half-space constraints simultaneously improves screening performance and speeds up lasso solvers.
When using the software, or parts of it, please cite: to appear
- Method: Seunghak Lee, Nico Goernitz, Eric P. Xing, David Heckerman, Christoph Lippert
- Software: Nico Goernitz & Seunghak Lee