Reweighing

Applies the reweighing algorithm to the dataset.

Inputs

• Data: reference dataset

Outputs

• Preprocessed Data: reference dataset with a meta-attribute weights added to it.

• Preprocessor: a preprocessor trained on the reference dataset.

Reweighing is a widget that mitigates bias in a dataset by assigning weights to individual instances in a way that encourages the model to prioritize learning from underrepresented groups while de-emphasizing overrepresented groups. This widget can be used in two ways:

• When the user provides the data as an input this widget will apply the reweighing algorithm to the dataset and output the preprocessed dataset with the meta-attribute weights added to it. The user can then use the preprocessed dataset as an input to other widgets. It will also output a preprocessor that can be used to apply the same transformation to a subset of the dataset.

• This widget can also be provided as an input to a learner widget. In this case the widget will be applied to any of the training datasets that are provided as an input to the learner widget.

Example

The first example shows how the Reweighing widget can be used to preprocess a dataset. First load a fairness dataset, in this case we will use the compas analysis dataset. We than split the dataset into a training and testing set using the Data Sampler widget. We connect the training set to the Reweighing widget which will train the algorithm and create a preprocessor. The preprocessor can be connected to the Apply Domain widget along with the testing set to apply the same transformation to the testing set. The preprocessed testing set can then be connected to the Dataset Bias widget to evaluate the bias of the dataset.

The second example demonstrates how to use the Reweighing widget as a preprocessor for a learner widget. We use it by connecting it and any other preprocessors we want to use into the Combine Preprocessors widget which we connect into the Weighted Logistic Regression widget. We then connect a dataset with fairness attributes and the learner into the Test & Score widget to evaluate the performance of the learner. In the evaluation results we can see the performance of the learner as well as the fairness metrics for its predictions.

Note, in this example we used the Weighted Logistic Regression and the Combine preprocessors widgets which are described in their respective sections.