Which is a better method Lasso regression or sequential Foward selection?

Question

I have a huge data of nearly 200 features and several of them might be correlated? In order to select best features which is preferred Lasso regression or sequential forward selection considering the correlation between features.(I can not run pca on this because feature transformation would transform features which cannot be interpreted).

• Few things I understood were: Lasso is really quick and sequential forward selection is very slow. But I am not sure how would lasso perform under correlated features i.e. I feel that the weights may be distributed among features(assume necessary features) which are correlated and because of low weights due to sharing Lasso might suggest they are irrelevant features.

Answer 1

I think there is no obvious advantage of one algorithm over another: I'd suggest running cross-validation to investigate which works best for the data at hand. You've suggested forward stepwise and Lasso, which both performs feature selection, you could also investigate backward stepwise and best-subset selection. Linear regression models are fast to train, so maybe you could investigate the cross-validation error for all 4 algorithms.

Hope this helps.