SLEP006: Metadata Routing

Author:

Joel Nothman, Adrin Jalali, Alex Gramfort, Thomas J. Fan

Status:

Accepted

Type:

Standards Track

Created:

2019-03-07

Abstract

This SLEP proposes an API to configure estimators, scorers, and CV splitters to request metadata when calling methods such as fit, predict, etc. Meta-estimators or functions that wrap estimators, scorers, or CV splitters will use this API to pass in the requested metadata.

Motivation and Scope

Scikit-learn has limited support for passing around information that is not (X, y). For example, to pass sample_weight to a step of a Pipeline, one needs to specify the step using dunder (__) prefixing:

>>> pipe = Pipeline([..., ('clf', LogisticRegression())])
>>> pipe.fit(X, y, clf__sample_weight=sample_weight)

Several other meta-estimators, such as GridSearchCV, support forwarding these fit parameters to their base estimator when fitting. Yet a number of important use cases are currently not supported:

  • Passing metadata (e.g. sample_weight) to a scorer used in cross-validation

  • Passing metadata (e.g. groups) to a CV splitter in nested cross-validation

  • Passing metadata (e.g. sample_weight) to some scorers and not others in multi-metric cross-validation. This is also required to handle fairness related metrics which usually expect one or more sensitive attributes to be passed to them along with the data.

  • Passing metadata to non-fit methods. For example, passing group indices for samples that are to be treated as a single sequence in prediction, or passing sensitive attributes to predict or transform of a fairness related estimator.

We define the following terms in this proposal:

  • consumer: An object that receives and consumes metadata, such as estimators, scorers, or CV splitters.

  • router: An object that passes metadata to a consumer or another router. Examples of routers include meta-estimators or functions. (For example GridSearchCV or cross_validate route sample weights, cross validation groups, etc. to consumers)

This SLEP proposes to add

  • get_metadata_routing to all consumers and routers (i.e. all estimators, scorers, and splitters supporting this API)

  • set_*_request to consumers (including estimators, scorers, and CV splitters), where * is a method that requires metadata. (e.g. set_fit_request, set_score_request, set_transform_request, etc.)

For example, set_fit_request configures an estimator to request metadata:

>>> log_reg = LogisticRegression().set_fit_request(sample_weight=True)

get_metadata_routing are used by routers to inspect the metadata needed by consumers. get_metadata_routing returns a MetadataRouter or a MetadataRequest object that stores and handles metadata routing. get_metadata_routing returns enough information for a router to know what metadata is requested, and whether the metadata is sample aligned or not. See the draft implementation for more implementation details.

Note that in the core library nothing is requested by default, except groups in Group*CV objects which request the groups metadata. At the time of writing this proposal, all metadata requested in the core library are sample aligned.

Also note that X, y, and Y input arguments are never automatically added to the routing mechanism and are always passed into their respective methods.

Detailed description

This SLEP unlocks many machine learning use cases that were not possible before. In this section, we will focus on some workflows that are made possible by this SLEP.

Nested Grouped Cross Validation

The following examples demonstrates nested grouped cross validation where a scorer and an estimator requests sample_weight and GroupKFold requests groups by default:

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

To support unweighted fitting and weighted scoring, metadata is set to False in fit_request:

>>> log_reg = (LogisticRegressionCV(cv=group_cv, scoring=weighted_acc)
...           .fit_request(sample_weight=False))
>>> cross_validate(
...     log_reg, X, y,
...     cv=GroupKFold(),
...     metadata={'sample_weight': weights, 'groups': groups},
...     scoring=weighted_acc)

Unweighted Feature selection

Consumers that do not accept weights during fitting such as SelectKBest will _not_ be routed weights:

>>> log_reg = (LogisticRegressionCV(cv=GroupKFold(), scoring=weighted_acc)
...            .set_fit_request(sample_weight=True))
>>> sel = SelectKBest(k=2)
>>> pipe = make_pipeline(sel, log_reg)
>>> pipe.fit(X, y, sample_weight=weights, groups=groups)

Note that if a consumer or a router starts accepting and consuming a certain metadata, the developer API enables developers to raise a warning and avoid silent behavior changes in users’ code. See the draft implementation for more details.

Different Scoring and Fitting Weights

We can pass different weights for scoring and fitting by using an aliases. In this example, scoring_weight is passed to the scoring and fitting_weight is passed to LogisticRegressionCV:

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

Nested Grouped Cross Validation with SearchCV

Since GroupKFold requests group metadata by default, GroupKFold instances can be passed to multiple routers to enable nested grouped cross validation. In this example, both RandomizedSearchCV and cross_validate set cv=GroupKFold() which enables grouped CV in the outer loop (cross_validate) and the inner random search:

>>> log_reg = LogisticRegression()
>>> distributions = {"C": uniform(loc=0, scale=4),
...                  "penalty": ['l2', 'l1']}
>>> random_search = RandomizedSearchCV(log_reg, distributions, cv=GroupKFold())
>>> cv_results = cross_validate(
...     log_reg, X, y,
...     cv=GroupKFold(),
...     metadata={"groups": my_groups})

Implementation

This SLEP has a draft implementation at #22083 by @adrinjalali. The implementation provides developer utilities that are used by scikit-learn and available to third-party estimators for adopting this SLEP. Specifically, the draft implementation makes it easier to define get_metadata_routing and set_*_request for consumers and routers.

Backward compatibility

Scikit-learn’s meta-estimators will deprecate the dunder (__) syntax for routing and enforce explicit request method calls. During the deprecation period, using dunder syntax routing and explicit request calls together will raise an error.

During the deprecation period, meta-estimators such as GridSearchCV will route fit_params to the inner estimators’ fit by default, but a deprecation warning is raised:

>>> # Deprecation warning, stating that the provided metadata is not requested
>>> GridSearchCV(LogisticRegression(), ...).fit(X, y, sample_weight=sw)

To avoid the warning, one would need to specify the request in LogisticRegression:

>>> grid = GridSearchCV(
...     LogisticRegression().set_fit_request(sample_weight=True), ...
... )
>>> grid.fit(X, y, sample_weight=sw)

Meta-estimators such as GridSearchCV will check which metadata is requested, and will error when metadata is passed in and the inner estimator is not configured to request it:

>>> weighted_acc = make_scorer(accuracy_score).score_request(sample_weight=True)
>>> log_reg = LogisticRegression()
>>> grid = GridSearchCV(log_reg, ..., scoring=weighted_scorer)
>>>
>>> # Raise a TypeError that log_reg is not specified with any routing
>>> # metadata for `sample_weight`, but sample_weight has been passed in to
>>> # `grid.fit`.
>>> grid.fit(X, y, sample_weight=sw)

To avoid the error, LogisticRegression must specify its metadata request by calling set_fit_request:

>>> # Request sample weights
>>> log_reg_weights = LogisticRegression().set_fit_request(sample_weight=True)
>>> grid = GridSearchCV(log_reg_with_weights, ...)
>>> grid.fit(X, y, sample_weight=sw)
>>>
>>> # Do not request sample_weights
>>> log_reg_no_weights = LogisticRegression().set_fit_request(sample_weight=False)
>>> grid = GridSearchCV(log_reg_no_weights, ...)
>>> grid.fit(X, y, sample_weight=sw)

Note that a meta-estimator will raise an error if the user passes a metadata which is not requested by any of the child objects of the meta-estimator.

Third-party estimators will need to adopt this SLEP in order to support metadata routing, while the dunder syntax is deprecated. Our implementation will provide developer APIs to trigger warnings and errors as described above to help with adopting this SLEP.

Alternatives

Over the years, there have been many proposed alternatives before we landed on this SLEP:

References and Footnotes