How to enforce monotonicity in FLAML? about flaml HOT 6 CLOSED

What's the definition of monotonicity? Assuming the data has only one feature x and requiring the model prediction as a monotonic function of x?

from flaml.

Yes. Ideally it would work as the monotone_constraints in xgboost.

from flaml.

It's easy to add your custom XGBoost learner by inheriting the class XGBoostEstimator (regression) or XGBoostSklearnEstimator (classification) from flaml.model. See section 3 of the notebook as an example.

In the constructor of the inherited estimator, set the monotonicity constraint you want to use in self.params,
after calling super().__init__(). For example, for regression, use

class MonotonicXGBoostEstimator(XGBoostEstimator):

----def __init__(self, **params):
--------super().__init__(**params)
--------self.params["monotone_constraints"] = "(1,-1)"

from flaml.

Thank you for the response!

One extra question: it seems that the goal of automl is to fit the best ML model rather than mere hyperparameter tuning, is there any plan to also test and return ensembles of base models built during the search (such as a linear ensemble of the best model of each algorithm) ?

Best,

from flaml.

You can set ensemble=True in the fit() function. It will build a stacked ensemble using the best models.

Please let me know if that works for you. Thanks.

from flaml.

It does. Thank you! For other questions, I will open a separate issue.

from flaml.