Open owenodriscoll opened 1 year ago
Perhaps by adding a method which not only selects best regressors but the best single regressor plus N
random other regressors in:
https://github.com/owenodriscoll/AutoML/blob/b2a83c5760903e85721d9e6cfdf98cae2fe4590d/AutoML/automl.py#L423
Add
n
)
instead of dedicating significant resources towards hyper-optimizing each regressor, train several of the same regressors (e.g. not lightgbm, xgboost, etc but lightgbm_0, lightgbm_1, lightgbm_2, xgboost_0, xgboost_1, xgboost_2 etc) with less resources. Thereby the stacking regressor has more information available by using more weak predictors which, though individually worse might contain more unique information, rather than several good predictors, which likely contain duplicate information.