Aki720
summary
Claims a concrete prior that outperforms Lasso on sparse high P low N regression, shows regularized horseshoe formulation that prevents complete separation in logistic models and keeps large signals unshrunk, gives Stan model code and diagnostics, reports experiments up to tens of thousands of features with thousands of observations, discusses computational cost and when dimensional reduction is required, and demonstrates that regularization reduces divergences and improves variable selection.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:30Motivation and context
- 3:56Sparse priors
- 6:08Understanding shrinkage
- 9:23The horseshoe prior in depth
- 14:37Regularized horseshoe
- 17:41Performance and results
- 19:10Q&A and conference talk
- 22:14Technical Q&A
related talks
44mActive Regression via Linear-Sample Sparsification
49mFast Algorithms for Regression Problems
17mUsing structure to select features in high dimension
09mOn the Power of Preconditioning in Sparse Linear Regression
02h 10mLasso and Ridge regression models in R: Introduction
45mHigh Dimensional Robust Sparse Regression
43mRobustness in High-Dimensional Inference Tasks
29mImplicit Regularization in Nonconvex Statistical Estimation