Tuning DataOps with Optuna

This example shows how to use Optuna to tune the hyperparameters of a skrub DataOp. As seen in the previous example, skrub DataOps can contain “choices”, objects created with choose_from(), choose_int(), choose_float(), etc. and we can use hyperparameter search techniques to pick the best outcome for each choice. Performing this search with Optuna allows us to benefit from its many features, such as state-of-the-art search strategies, monitoring and visualization, stopping and resuming searches, and parallel or distributed computation.

In order to use Optuna with skrub, the package must be installed first. This can be done with pip:

pip install optuna

A simple regressor and example data.

We will fit a regressor containing a few choices on a toy dataset. We try 2 regressors: extra trees and ridge. They both have hyperparameters that we want to tune.

from sklearn.ensemble import ExtraTreesRegressor
from sklearn.linear_model import Ridge

import skrub

extra_tree = ExtraTreesRegressor(
    min_samples_leaf=skrub.choose_int(1, 32, log=True, name="min_samples_leaf"),
)
ridge = Ridge(alpha=skrub.choose_float(0.01, 10.0, log=True, name="α"))

regressor = skrub.choose_from(
    {"extra_tree": extra_tree, "ridge": ridge}, name="regressor"
)
data = skrub.var("data")
X = data.drop(columns="MedHouseVal", errors="ignore").skb.mark_as_X()
y = data["MedHouseVal"].skb.mark_as_y()
pred = X.skb.apply(regressor, y=y)
print(pred.skb.describe_param_grid())

- regressor: 'extra_tree'
  min_samples_leaf: choose_int(1, 32, log=True, name='min_samples_leaf')
- regressor: 'ridge'
  α: choose_float(0.01, 10.0, log=True, name='α')

Load data for the example

from sklearn.model_selection import KFold

# (We subsample the dataset by half to make the example run faster)
df = skrub.datasets.fetch_california_housing().california_housing.sample(
    10_000, random_state=0
)

# The environment we will use to fit the learners created by our DataOp.
env = {"data": df}
cv = KFold(n_splits=4, shuffle=True, random_state=0)

Selecting the best hyperparameters with Optuna.

The simplest way to use Optuna is to pass backend='optuna' to DataOp.skb.make_randomized_search(). It is used very similarly as with the default backend (sklearn.model_selection.RandomizedSearchCV). Additional parameters are available to control the Optuna sampler, storage and study name, and timeout. Note that in order to persist the study and resume it later, the storage parameter must be set to a valid database URL (e.g., a SQLite file). Refer to the User Guide for an example.

search = pred.skb.make_randomized_search(
    backend="optuna", cv=cv, n_iter=10, random_state=10
)
search.fit(env)
search.results_

Running optuna search for study skrub_randomized_search_9308ffce-a740-429b-9506-1ce0c8c8f488 in storage /tmp/tmpw_twijg1_skrub_optuna_search_storage/optuna_storage
[I 2026-02-10 10:24:59,185] A new study created in Journal with name: skrub_randomized_search_9308ffce-a740-429b-9506-1ce0c8c8f488
[I 2026-02-10 10:24:59,426] Trial 0 finished with value: 0.5998958893326463 and parameters: {'2:regressor': '1:ridge', '1:α': 0.08737056333717456}. Best is trial 0 with value: 0.5998958893326463.
[I 2026-02-10 10:24:59,658] Trial 1 finished with value: 0.5998965729709885 and parameters: {'2:regressor': '1:ridge', '1:α': 0.17777676175509427}. Best is trial 1 with value: 0.5998965729709885.
[I 2026-02-10 10:24:59,897] Trial 2 finished with value: 0.5998956300607201 and parameters: {'2:regressor': '1:ridge', '1:α': 0.05323908243249831}. Best is trial 1 with value: 0.5998965729709885.
[I 2026-02-10 10:25:00,128] Trial 3 finished with value: 0.5998953213503151 and parameters: {'2:regressor': '1:ridge', '1:α': 0.012709575518456051}. Best is trial 1 with value: 0.5998965729709885.
[I 2026-02-10 10:25:05,181] Trial 4 finished with value: 0.7897936635651363 and parameters: {'2:regressor': '0:extra_tree', '0:min_samples_leaf': 2}. Best is trial 4 with value: 0.7897936635651363.
[I 2026-02-10 10:25:13,174] Trial 5 finished with value: 0.7930210657931196 and parameters: {'2:regressor': '0:extra_tree', '0:min_samples_leaf': 1}. Best is trial 5 with value: 0.7930210657931196.
[I 2026-02-10 10:25:13,404] Trial 6 finished with value: 0.5998987110257884 and parameters: {'2:regressor': '1:ridge', '1:α': 0.464456825115831}. Best is trial 5 with value: 0.7930210657931196.
[I 2026-02-10 10:25:16,241] Trial 7 finished with value: 0.7586118873804404 and parameters: {'2:regressor': '0:extra_tree', '0:min_samples_leaf': 7}. Best is trial 5 with value: 0.7930210657931196.
[I 2026-02-10 10:25:19,752] Trial 8 finished with value: 0.7745354458535161 and parameters: {'2:regressor': '0:extra_tree', '0:min_samples_leaf': 4}. Best is trial 5 with value: 0.7930210657931196.
[I 2026-02-10 10:25:24,730] Trial 9 finished with value: 0.7910036605628203 and parameters: {'2:regressor': '0:extra_tree', '0:min_samples_leaf': 2}. Best is trial 5 with value: 0.7930210657931196.
[I 2026-02-10 10:25:24,732] A new study created in memory with name: skrub_randomized_search_9308ffce-a740-429b-9506-1ce0c8c8f488

	min_samples_leaf	α	regressor	mean_test_score
0	1.0	NaN	extra_tree	0.793021
1	2.0	NaN	extra_tree	0.791004
2	2.0	NaN	extra_tree	0.789794
3	4.0	NaN	extra_tree	0.774535
4	7.0	NaN	extra_tree	0.758612
5	NaN	0.464457	ridge	0.599899
6	NaN	0.177777	ridge	0.599897
7	NaN	0.087371	ridge	0.599896
8	NaN	0.053239	ridge	0.599896
9	NaN	0.012710	ridge	0.599895

The usual results_, detailed_results_ and plot_results() are still available.

search.plot_results()

The Optuna Study that was used to run the hyperparameter search is available in the attribute study_:

search.study_

<optuna.study.study.Study object at 0x7f9c70425e00>

search.study_.best_params

{'2:regressor': '0:extra_tree', '0:min_samples_leaf': 1}

This allows us to use Optuna’s reporting capabilities provided in optuna.visualization or optuna-dashboard.

import optuna

optuna.visualization.plot_slice(search.study_, params=["0:min_samples_leaf"])

Using Optuna directly for more advanced use cases

Often we may want more control over the use of Optuna, or to access functionality not available through DataOp.skb.make_randomized_search() such as the ask-and-tell interface, trial pruning, callbacks, multi-objective optimization, etc. .

Directly using Optuna ourselves is also easy, as we will show now. What makes this possible is that we can pass an Optuna Trial to DataOp.skb.make_learner() in which case the parameters suggested by the trial are used to create the learner.

We revisit the example above, following the typical Optuna workflow.

The optuna.Study runs the hyperparameter search.

Its method optimize is given an objective function. The objective must accept a Trial object (which is produced by the study and picks the parameters for a given evaluation of the objective) and return the value to maximize (or minimize).

To use Optuna with a DataOp, we just need to pass the Trial object to DataOp.skb.make_learner(). This creates a SkrubLearner initialized with the parameters picked by the optuna Trial.

We can then cross-validate the SkrubLearner, or score it however we prefer, and return the score so that the optuna Study can take it into account.

Here we return a single score (R²), but multi-objective optimization is also possible. Please refer to the Optuna documentation for more information.

def objective(trial):
    learner = pred.skb.make_learner(choose=trial)
    cv_results = skrub.cross_validate(learner, environment=env, cv=cv)
    return cv_results["test_score"].mean()


study = optuna.create_study(direction="maximize")
study.optimize(objective, n_trials=10)
study.best_params

[I 2026-02-10 10:25:30,185] A new study created in memory with name: no-name-d2c3799e-f1f7-475f-9ab5-fc8eb285afc5
[I 2026-02-10 10:25:30,415] Trial 0 finished with value: 0.5999388375487643 and parameters: {'2:regressor': '1:ridge', '1:α': 7.778421443913227}. Best is trial 0 with value: 0.5999388375487643.
[I 2026-02-10 10:25:38,452] Trial 1 finished with value: 0.7921475110890374 and parameters: {'2:regressor': '0:extra_tree', '0:min_samples_leaf': 1}. Best is trial 1 with value: 0.7921475110890374.
[I 2026-02-10 10:25:38,665] Trial 2 finished with value: 0.5998968471096364 and parameters: {'2:regressor': '1:ridge', '1:α': 0.21419820409047396}. Best is trial 1 with value: 0.7921475110890374.
[I 2026-02-10 10:25:43,628] Trial 3 finished with value: 0.791341354843085 and parameters: {'2:regressor': '0:extra_tree', '0:min_samples_leaf': 2}. Best is trial 1 with value: 0.7921475110890374.
[I 2026-02-10 10:25:43,839] Trial 4 finished with value: 0.5998984275446252 and parameters: {'2:regressor': '1:ridge', '1:α': 0.42609550480521413}. Best is trial 1 with value: 0.7921475110890374.
[I 2026-02-10 10:25:51,816] Trial 5 finished with value: 0.795523106005046 and parameters: {'2:regressor': '0:extra_tree', '0:min_samples_leaf': 1}. Best is trial 5 with value: 0.795523106005046.
[I 2026-02-10 10:25:52,026] Trial 6 finished with value: 0.5998955359166415 and parameters: {'2:regressor': '1:ridge', '1:α': 0.040866575407271455}. Best is trial 5 with value: 0.795523106005046.
[I 2026-02-10 10:25:56,040] Trial 7 finished with value: 0.7813008825764698 and parameters: {'2:regressor': '0:extra_tree', '0:min_samples_leaf': 3}. Best is trial 5 with value: 0.795523106005046.
[I 2026-02-10 10:25:56,252] Trial 8 finished with value: 0.5999058083279392 and parameters: {'2:regressor': '1:ridge', '1:α': 1.462978585285932}. Best is trial 5 with value: 0.795523106005046.
[I 2026-02-10 10:25:58,365] Trial 9 finished with value: 0.7201796077975192 and parameters: {'2:regressor': '0:extra_tree', '0:min_samples_leaf': 19}. Best is trial 5 with value: 0.795523106005046.

{'2:regressor': '0:extra_tree', '0:min_samples_leaf': 1}

We can also use Optuna’s visualization capabilities to inspect the study:

optuna.visualization.plot_optimization_history(study)

Now we build a learner with the best hyperparameters and fit it on the full dataset:

best_learner = pred.skb.make_learner(choose=study.best_trial)

# This would achieve the same result:
# best_learner = pred.skb.make_learner()
# best_learner.set_params(**study.best_params)

best_learner.fit(env)
print(best_learner.describe_params())

{'min_samples_leaf': 1, 'regressor': 'extra_tree'}

Estimated memory usage: 613 MB