|
Abstract:
|
Enforced sparsity for Bayesian spike and slab models is a new regularization and variable selection method. Enforced sparsity shrinks regression coefficients exactly to zero and can be implemented using standard Gibbs sampling methods. The idea is motivated by theory for selective shrinkage; a unique property whereby the posterior mean is shrunk towards zero for non-informative variables. Using real and synthetic data, I compare enforced spike and slab sparsity to the lasso, random forests, and boosting and find it yields sparse and accurate models, and performs well even when the number of variables is larger than the sample size.
|
