How does one use the Fairlearn metrics to make a decision on whether a feature is biased or not?

Question

Can anyone suggest best practice guidelines on selecting thresholds for the disparity metrics to determine if a sensitive attribute is biased or not?

Answer 1

This is a great question! To figure this out, it really helps to frame the impact of these analytical choices in terms of how it measures real harms to real people.

If you're into research papers, Language (Technology) is Power: A Critical Survey of “Bias” in NLP (Blodget et al. 2020) suggests some ways to approach fairness work that can lead to more productive conversations than focusing on bias alone.

More practically, framing the conversation in terms of real harms to real people then allows you to express the impact of the threshold choices for fairness metrics in accessible human terms. That can go along way to illustrative to various stakeholders why this work matters and is worth doing.

To sketch this out a bit more, false positives often lead to different harms than false negatives, and if there are human review processes this influences how you might quantify these kinds of harms or risks. Upstream labeling noise can influence how much trust you put in the thresholding procedure to capture real harms. For decision support systems, downstream engagement, adoption, and trust in predictive systems often influences whether human decision makers actually make use of model predictions. A few lines of code can show stakeholders the impact of that kind of upstream or downstream noise on fairness metrics, and show stakeholders other ways that the technical system may be amplifying real harms.

If you want to chat more, or dig into more specifics that would help you explore this or kick off those conversation on your team, feel free to ask in https://gitter.im/fairlearn/community as well. Like most software engineering work, it's easier to give more actionable suggestions within a specific application, context or set of constraints.