
You are working on a supervised learning model and have more candidate features than you want to ship. Some are noisy, some are correlated, and some may only help in certain slices of the data.
How do you approach feature selection for a machine learning model?