AI Fairness 360

The IBM AI Fairness 360 Toolkit contains several bias mitigation algorithms that are applicable to various stages of the machine learning pipeline. The toolkit implements different notions of fairness, both on individual and the group level, and several fairness metrics for both classes of fairness. The toolkit provides additional guidance on choosing metrics and mitigation algorithms given a particular goal and application.

The following should be noted when using the fairness toolkit (and other similar toolkits, for that matter):

The toolkit should only be used in a very limited setting: allocation or risk assessment problems with well-defined protected attributes in which one would like to have some sort of statistical or mathematical notion of sameness. Even then, the code and collateral contained in AIF360 is only a starting point to a broader discussion among multiple stakeholders on overall decision making workflows. source

Moreover, the choice for a particular algorithm from the toolkit also depends on assumptions on the equality of people. Within the group fairness approach, the toolkit distinguishes two different “worldviews” underlying group fairness: the “we’re all equal” worldview and the “what you see if what you get worldview”. See group fairness for a further explanation.

To choose an approach, we need to consider 1) the type of fairness to strive for 2) given that type, which fairness metrics/constraints to use (if applicable) 3) where in the pipeline to intervene 4) which algorithm is suitable to support the made choices.

Choosing fairness metrics

The amount of implemented fairness metrics is very large, so we refer here to the API and summarize the directions given in the guidance material.

Individual vs. group

For individual fairness, refer to the SampleDistortionMetric.

For group fairness, refer to DatasetMetric and its children classes BinaryLabelDatasetMetric and ClassificationMetric.

It is possible to combine individual- and group fairness in a single metric. The ClassificationMetric contains several measures related to the generalized entropy index suitable for this purpose.

Stages

The metrics under (BinaryLabel)DatasetMetric apply to the training data and are thus relevant for the preprocessing stage.

The metrics under ClassificationMetric apply to the models themselves and are thus relevant during the in-processing stage.

Worldviews

In the case of group fairness, the underlying “worldview” can be a relevant factor in choosing metrics.

For “we are all equal” worldview the demographic parity metrics are appropriate, e.g.

  • disparate_impact
  • statistical_parity_difference

If the application follows the “what you see is what you get” worldview, then variations of the equality of odds fairness metric is appropriate:

  • average_odds_difference
  • average_abs_odds_difference

Some metrics are not specific to a particular worldview, such as metrics based on error rates, e.g.:

  • false_negative_rate_ratio
  • false_positive_rate_ratio
  • error_rate_ratio

Choosing algorithms

The API nicely lists algorithms by the stage they are applicable in. The algorithms that affect the in-processing stage are mostly suitable for classification. However, a reduction-based approach such as GridSearchReduction can also be used for regression.

Preprocessing

For preprocessing algorithms see aif360.algorithms.preprocessing:

Relevant remarks from the guidance material on the preprocessing algorithms:

  • “Among pre-processing algorithms, reweighing only changes weights applied to training samples; it does not change any feature or label values. Therefore, it may be a preferred option in case the application does not allow for value changes.”
  • “Disparate impact remover and optimized pre-processing yield modified datasets in the same space as the input training data, whereas LFR’s pre-processed dataset is in a latent space.”
  • “If the application requires transparency on the transformation, then disparate impact remover and optimized pre-processing may be preferred options”
  • “Moreover, optimized pre-processing addresses both group fairness and individual fairness.”

In-processing

For algorithms for fair learning see aif360.algorithms.inprocessing.

Relevant remarks from the guidance material on the “in-processing” algorithms:

  • “Among in-processing algorithms, the prejudice remover is limited to learning algorithms that allow for regularization terms whereas the adversarial debiasing algorithm allows for a more general set of learning algorithms, and may be preferred for that reason.”

Post-processing

For post-processing algorithms, see aif360.algorithms.postprocessing.

Relevant remarks from the guidance material on the “post-processing” algorithms:

  • “Among post-processing algorithms, the two equalized odds post-processing algorithms have a randomized component whereas the reject option algorithm is deterministic, and may be preferred for that reason.”