fairlearn.preprocessing.CorrelationRemover#

class fairlearn.preprocessing.CorrelationRemover(*, sensitive_feature_ids=(), alpha=1)[source]#

A component that filters out sensitive correlations in a dataset.

CorrelationRemover applies a linear transformation to the non-sensitive feature columns in order to remove their correlation with the sensitive feature columns while retaining as much information as possible (as measured by the least-squares error).

Read more in the User Guide.

Parameters:
sensitive_feature_idslist

list of columns to filter out this can be a sequence of either int, in the case of numpy, or string, in the case of pandas.

alphafloat

parameter to control how much to filter, for alpha=1.0 we filter out all information while for alpha=0.0 we don’t apply any.

Notes

This method will change the original dataset by removing all correlation with sensitive values. In mathematical terms, assume we have the original dataset \(\mathbf{X}\), which contains a set of sensitive features denoted by \(\mathbf{S}\) and a set of non-sensitive features denoted by \(\mathbf{Z}\). The goal is to remove correlations between the sensitive features and the non-sensitive features.

Let \(m_s\) and \(m_{ns}\) denote the number of sensitive and non-sensitive features, respectively. Let \(\bar{\mathbf{s}}\) represent the mean of the sensitive features, i.e., \(\bar{\mathbf{s}} = [\bar{s}_1, \dots, \bar{s}_{m_s}]^\top\), where \(\bar{s}_j\) is the mean of the \(j\text{-th}\) sensitive feature.

For each non-sensitive feature \(\mathbf{z}_j\in\mathbb{R}^n\), where \(j=1,\dotsc,m_{ns}\), we compute an optimal weight vector \(\mathbf{w}_j^* \in \mathbb{R}^{m_s}\) that minimizes the following least squares objective:

\[\min _{\mathbf{w}} \| \mathbf{z}_j - (\mathbf{S}-\mathbf{1}_n\times\bar{\mathbf{s}}^\top) \mathbf{w} \|_2^2\]

where \(\mathbf{1}_n\) is the all-one vector in \(\mathbb{R}^n\).

In other words, \(\mathbf{w}_j^*\) is the solution to a linear regression problem where we project \(\mathbf{z}_j\) onto the centered sensitive features. The weight matrix \(\mathbf{W}^* = (\mathbf{w}_1^*, \dots, \mathbf{w}_{m_{ns}}^*)\) is thus obtained by solving this regression for each non-sensitive feature.

Once we have the optimal weight matrix \(\mathbf{W}^*\), we compute the residual non-sensitive features \(\mathbf{Z}^*\) as follows:

\[\mathbf{Z}^* = \mathbf{Z} - (\mathbf{S}-\mathbf{1}_n\times\bar{\mathbf{s}}^\top) \mathbf{W}^*\]

The columns in \(\mathbf{S}\) will be dropped from the dataset \(\mathbf{X}\), and \(\mathbf{Z}^*\) will replace the original non-sensitive features \(\mathbf{Z}\), but the hyper parameter \(\alpha\) does allow you to tweak the amount of filtering that gets applied:

\[\mathbf{X}_{\text{tfm}} = \alpha \mathbf{X}_{\text{filtered}} + (1-\alpha) \mathbf{X}_{\text{orig}}\]

Note that the lack of correlation does not imply anything about statistical dependence. Therefore, we expect this to be most appropriate as a preprocessing step for (generalized) linear models.

New in version 0.6.

fit(X, y=None)[source]#

Learn the projection required to make the dataset uncorrelated with sensitive columns.

fit_transform(X, y=None, **fit_params)[source]#

Fit to data, then transform it.

Fits transformer to X and y with optional parameters fit_params and returns a transformed version of X.

Parameters:
Xarray-like of shape (n_samples, n_features)

Input samples.

yarray-like of shape (n_samples,) or (n_samples, n_outputs), default=None

Target values (None for unsupervised transformations).

**fit_paramsdict

Additional fit parameters.

Returns:
X_newndarray array of shape (n_samples, n_features_new)

Transformed array.

get_metadata_routing()[source]#

Get metadata routing of this object.

Please check User Guide on how the routing mechanism works.

Returns:
routingMetadataRequest

A MetadataRequest encapsulating routing information.

get_params(deep=True)[source]#

Get parameters for this estimator.

Parameters:
deepbool, default=True

If True, will return the parameters for this estimator and contained subobjects that are estimators.

Returns:
paramsdict

Parameter names mapped to their values.

set_output(*, transform=None)[source]#

Set output container.

See Introducing the set_output API for an example on how to use the API.

Parameters:
transform{“default”, “pandas”, “polars”}, default=None

Configure output of transform and fit_transform.

  • “default”: Default output format of a transformer

  • “pandas”: DataFrame output

  • “polars”: Polars output

  • None: Transform configuration is unchanged

New in version 1.4: “polars” option was added.

Returns:
selfestimator instance

Estimator instance.

set_params(**params)[source]#

Set the parameters of this estimator.

The method works on simple estimators as well as on nested objects (such as Pipeline). The latter have parameters of the form <component>__<parameter> so that it’s possible to update each component of a nested object.

Parameters:
**paramsdict

Estimator parameters.

Returns:
selfestimator instance

Estimator instance.

transform(X)[source]#

Transform X by applying the correlation remover.