Fairness evaluation

[Jeanselme]

Here are some thoughts concerning fairness evaluation:

Protected attributes

Extraction: As methodologies may implement pre-, in- or post-processing to enhance algorithmic fairness, I think we need to extract sensitive attributes at the level of task extraction, not only at evaluation.

Definition: Sensitive attributes may differ across datasets (for instance, ethnicity changes for a patient at different visits in the MIMIC dataset; Canada and France do not record ethnicity; and different datasets may present different levels of granularity). As an initial step, we could focus on age and sex.

Save: Attributes could be saved in the target file and extracted at the same time. Alternatively, we could gather them in a separate file, following a different script.

Metrics

We can distinguish three fairness definitions:
Group fairness: A model is fair if performance is equal across groups defined by protected attributes.
Causal fairness: A model is fair if the prediction remains unchanged if membership changes.
Individual fairness: A model is fair if similar individuals are treated similarly.

We need the causal graph to estimate causal fairness and a meaningful distance for individual fairness. I think, as a first step, group fairness is the simplest to implement (and is widely used in medical ML). So, we could stratify performance per identified protected groups and compute the difference as a measure of fairness.

Let me know what you think!

[Jeanselme]

Adding @kamilest

[kamilest]

• Update dataset.md #31 The "new dataset" template should request fairness-related information and relevant/supported group definitions.

[Jeanselme]

Following our conversation, we could start with the following fairness metrics for binary tasks:

• Group specific AUC (just stratifying given the different groups)
• Demographic parity: measure if the expected rate are the same across groups
• Equalized Odds: same but conditional on outcome