I'm in the feature filtering process stage. May I ask what are the differences and advantages and disadvantages of the two methods of filtering features, adding features one by one and observing the change in RMSE and removing features one by one from the tail?
I first filtered out the features with correlation of 0.6 or less with the target value and then trained a feature importance with lightgbm, so which of the above two methods would be better? Gradually adding features to record rmse changes taking the smallest, or culling features one by one?