kyle-messier
commented
7 months ago @ericbair-sciome @sciome-bot In terms of implementing this - we can still use the multivariate Gaussian LASSO from glmnet. After the betas have been fit using lasso, just re-estimate using an un-penalized linear regression (e.g. glm or lm). This is equivalent to the relaxed-lasso with gamma = 0, i.e. the most relaxed :relaxed:
Kyle found that relaxed lasso worked better than regular lasso in some situations. This should be evaluated more carefully. If the relaxed lasso generally outperforms the regular lasso, then it should be used as the default.