Equivalent of MultiOutputRegressor with different hyperparameters for each output?

[eschibli]

I need a composite estimator for vector targets made up of individual estimators, similar to what is provided by MultiOutputRegressor. However, I need a model with different hyperparameters for each target.

My first instinct was to define a MultiOutputRegressor of dummy estimators, then overwrite the estimators_ attribute with my desired regressors, but this does not work as only the base estimator is defined on construction; it is then copied on fit.

Do I need to write my own meta-estimator class, or is there a better solution I'm not thinking of?

[glemaitre]

Yes, you will need to create a new class where you can pass a list of estimators (as many estimators as the number of targets).

However, if you plan to tune the hyperparameters, you can indeed pass a GridSearchCV as an estimator and then the hyperparameter will be tuned for each target independently.