1. Metadata Routing#

Note

The Metadata Routing API is experimental, and is not yet implemented for all estimators. Please refer to the list of supported and unsupported models for more information. It may change without the usual deprecation cycle. By default this feature is not enabled. You can enable it by setting the enable_metadata_routing flag to True:

>>> import sklearn
>>> sklearn.set_config(enable_metadata_routing=True)

Note that the methods and requirements introduced in this document are only relevant if you want to pass metadata (e.g. sample_weight) to a method. If you’re only passing X and y and no other parameter / metadata to methods such as fit, transform, etc., then you don’t need to set anything.

This guide demonstrates how metadata can be routed and passed between objects in scikit-learn. If you are developing a scikit-learn compatible estimator or meta-estimator, you can check our related developer guide: Metadata Routing.

Metadata is data that an estimator, scorer, or CV splitter takes into account if the user explicitly passes it as a parameter. For instance, KMeans accepts sample_weight in its fit() method and considers it to calculate its centroids. classes are consumed by some classifiers and groups are used in some splitters, but any data that is passed into an object’s methods apart from X and y can be considered as metadata. Prior to scikit-learn version 1.3, there was no single API for passing metadata like that if these objects were used in conjunction with other objects, e.g. a scorer accepting sample_weight inside a GridSearchCV.

With the Metadata Routing API, we can transfer metadata to estimators, scorers, and CV splitters using meta-estimators (such as Pipeline or GridSearchCV) or functions such as cross_validate which route data to other objects. In order to pass metadata to a method like fit or score, the object consuming the metadata, must request it. This is done via set_{method}_request() methods, where {method} is substituted by the name of the method that requests the metadata. For instance, estimators that use the metadata in their fit() method would use set_fit_request(), and scorers would use set_score_request(). These methods allow us to specify which metadata to request, for instance set_fit_request(sample_weight=True).

For grouped splitters such as GroupKFold, a groups parameter is requested by default. This is best demonstrated by the following examples.

1.1. Usage Examples#

Here we present a few examples to show some common use-cases. Our goal is to pass sample_weight and groups through cross_validate, which routes the metadata to LogisticRegressionCV and to a custom scorer made with make_scorer, both of which can use the metadata in their methods. In these examples we want to individually set whether to use the metadata within the different consumers.

The examples in this section require the following imports and data:

>>> import numpy as np
>>> from sklearn.metrics import make_scorer, accuracy_score
>>> from sklearn.linear_model import LogisticRegressionCV, LogisticRegression
>>> from sklearn.model_selection import cross_validate, GridSearchCV, GroupKFold
>>> from sklearn.feature_selection import SelectKBest
>>> from sklearn.pipeline import make_pipeline
>>> n_samples, n_features = 100, 4
>>> rng = np.random.RandomState(42)
>>> X = rng.rand(n_samples, n_features)
>>> y = rng.randint(0, 2, size=n_samples)
>>> my_groups = rng.randint(0, 10, size=n_samples)
>>> my_weights = rng.rand(n_samples)
>>> my_other_weights = rng.rand(n_samples)

1.1.1. Weighted scoring and fitting#

The splitter used internally in LogisticRegressionCV, GroupKFold, requests groups by default. However, we need to explicitly request sample_weight for it and for our custom scorer by specifying sample_weight=True in LogisticRegressionCV`s `set_fit_request() method and in make_scorer`s `set_score_request method. Both consumers know how to use sample_weight in their fit() or score() methods. We can then pass the metadata in cross_validate which will route it to any active consumers:

>>> weighted_acc = make_scorer(accuracy_score).set_score_request(sample_weight=True)
>>> lr = LogisticRegressionCV(
...     cv=GroupKFold(),
...     scoring=weighted_acc
... ).set_fit_request(sample_weight=True)
>>> cv_results = cross_validate(
...     lr,
...     X,
...     y,
...     params={"sample_weight": my_weights, "groups": my_groups},
...     cv=GroupKFold(),
...     scoring=weighted_acc,
... )

Note that in this example, cross_validate routes my_weights to both the scorer and LogisticRegressionCV.

If we would pass sample_weight in the params of cross_validate, but not set any object to request it, UnsetMetadataPassedError would be raised, hinting to us that we need to explicitly set where to route it. The same applies if params={"sample_weights": my_weights, ...} were passed (note the typo, i.e. weights instead of weight), since sample_weights was not requested by any of its underlying objects.

1.1.2. Weighted scoring and unweighted fitting#

When passing metadata such as sample_weight into a router (meta-estimators or routing function), all sample_weight consumers require weights to be either explicitly requested or explicitly not requested (i.e. True or False). Thus, to perform an unweighted fit, we need to configure LogisticRegressionCV to not request sample weights, so that cross_validate does not pass the weights along:

>>> weighted_acc = make_scorer(accuracy_score).set_score_request(sample_weight=True)
>>> lr = LogisticRegressionCV(
...     cv=GroupKFold(), scoring=weighted_acc,
... ).set_fit_request(sample_weight=False)
>>> cv_results = cross_validate(
...     lr,
...     X,
...     y,
...     cv=GroupKFold(),
...     params={"sample_weight": my_weights, "groups": my_groups},
...     scoring=weighted_acc,
... )

If linear_model.LogisticRegressionCV.set_fit_request had not been called, cross_validate would raise an error because sample_weight is passed but LogisticRegressionCV would not be explicitly configured to recognize the weights.

1.1.3. Unweighted feature selection#

Routing metadata is only possible if the object’s method knows how to use the metadata, which in most cases means they have it as an explicit parameter. Only then we can set request values for metadata using set_fit_request(sample_weight=True), for instance. This makes the object a consumer.

Unlike LogisticRegressionCV, SelectKBest can’t consume weights and therefore no request value for sample_weight on its instance is set and sample_weight is not routed to it:

>>> weighted_acc = make_scorer(accuracy_score).set_score_request(sample_weight=True)
>>> lr = LogisticRegressionCV(
...     cv=GroupKFold(), scoring=weighted_acc,
... ).set_fit_request(sample_weight=True)
>>> sel = SelectKBest(k=2)
>>> pipe = make_pipeline(sel, lr)
>>> cv_results = cross_validate(
...     pipe,
...     X,
...     y,
...     cv=GroupKFold(),
...     params={"sample_weight": my_weights, "groups": my_groups},
...     scoring=weighted_acc,
... )

1.1.4. Different scoring and fitting weights#

Despite make_scorer and LogisticRegressionCV both expecting the key sample_weight, we can use aliases to pass different weights to different consumers. In this example, we pass scoring_weight to the scorer, and fitting_weight to LogisticRegressionCV:

>>> weighted_acc = make_scorer(accuracy_score).set_score_request(
...    sample_weight="scoring_weight"
... )
>>> lr = LogisticRegressionCV(
...     cv=GroupKFold(), scoring=weighted_acc,
... ).set_fit_request(sample_weight="fitting_weight")
>>> cv_results = cross_validate(
...     lr,
...     X,
...     y,
...     cv=GroupKFold(),
...     params={
...         "scoring_weight": my_weights,
...         "fitting_weight": my_other_weights,
...         "groups": my_groups,
...     },
...     scoring=weighted_acc,
... )

1.2. API Interface#

A consumer is an object (estimator, meta-estimator, scorer, splitter) which accepts and uses some metadata in at least one of its methods (for instance fit, predict, inverse_transform, transform, score, split). Meta-estimators which only forward the metadata to other objects (child estimators, scorers, or splitters) and don’t use the metadata themselves are not consumers. (Meta-)Estimators which route metadata to other objects are routers. A(n) (meta-)estimator can be a consumer and a router at the same time. (Meta-)Estimators and splitters expose a set_{method}_request method for each method which accepts at least one metadata. For instance, if an estimator supports sample_weight in fit and score, it exposes estimator.set_fit_request(sample_weight=value) and estimator.set_score_request(sample_weight=value). Here value can be:

  • True: method requests a sample_weight. This means if the metadata is provided, it will be used, otherwise no error is raised.

  • False: method does not request a sample_weight.

  • None: router will raise an error if sample_weight is passed. This is in almost all cases the default value when an object is instantiated and ensures the user sets the metadata requests explicitly when a metadata is passed. The only exception are Group*Fold splitters.

  • "param_name": alias for sample_weight if we want to pass different weights to different consumers. If aliasing is used the meta-estimator should not forward "param_name" to the consumer, but sample_weight instead, because the consumer will expect a param called sample_weight. This means the mapping between the metadata required by the object, e.g. sample_weight and the variable name provided by the user, e.g. my_weights is done at the router level, and not by the consuming object itself.

Metadata are requested in the same way for scorers using set_score_request.

If a metadata, e.g. sample_weight, is passed by the user, the metadata request for all objects which potentially can consume sample_weight should be set by the user, otherwise an error is raised by the router object. For example, the following code raises an error, since it hasn’t been explicitly specified whether sample_weight should be passed to the estimator’s scorer or not:

>>> param_grid = {"C": [0.1, 1]}
>>> lr = LogisticRegression().set_fit_request(sample_weight=True)
>>> try:
...     GridSearchCV(
...         estimator=lr, param_grid=param_grid
...     ).fit(X, y, sample_weight=my_weights)
... except ValueError as e:
...     print(e)
[sample_weight] are passed but are not explicitly set as requested or not
requested for LogisticRegression.score, which is used within GridSearchCV.fit.
Call `LogisticRegression.set_score_request({metadata}=True/False)` for each metadata
you want to request/ignore.

The issue can be fixed by explicitly setting the request value:

>>> lr = LogisticRegression().set_fit_request(
...     sample_weight=True
... ).set_score_request(sample_weight=False)

At the end of the Usage Examples section, we disable the configuration flag for metadata routing:

>>> sklearn.set_config(enable_metadata_routing=False)

1.3. Metadata Routing Support Status#

All consumers (i.e. simple estimators which only consume metadata and don’t route them) support metadata routing, meaning they can be used inside meta-estimators which support metadata routing. However, development of support for metadata routing for meta-estimators is in progress, and here is a list of meta-estimators and tools which support and don’t yet support metadata routing.

Meta-estimators and functions supporting metadata routing:

Meta-estimators and tools not supporting metadata routing yet: