Data science without prejudice – How to avoid unfair conclusions even if they are true?” session leader dr. Maurits Kaptein. To future-proof responsible data science methods, foundational research is needed, and, given the complementarity of TU/e and TiU in JADS, there are great opportunities to collaborate on this theme. This was reflected by the JADS Workshop on Responsible Data Science

1. 2 Dec. 2017, JADS
Discussion on Fairness in Data Science
Moderated by Dr. Maurits C. Kaptein

2. Fairness: How to avoid unfair conclusions?
Accuracy: How do we guarantee accuracy?
Confidentiality: How to not reveal secrets?
Transparency: How to make answers indisputable?

3. First focus on Fairness; lets assume the model output is
accurate and transparent…

4. What is Fairness?
- Non discriminating?
- Broader definitions?
- Veil of ignorance?
Canonical example of unfair Data Science: Higher bail-out
charges for African Americans based on predicted risk of
recurrence
Search for examples….

5. Fair? (show of hands)
Decision to search a person in the airport based on
membership of a terrorist group.

6. Fair?
Decision to search a person in the airport based on
religious beliefs.

7. Fair?
Decision to search a person in the airport based on a set of
features that does not include religious belief, but does
predict it accurately.

8. Fair?
Decision to search a person in the airport based on race.

9. Fairness implemented:
- Is a distinct set of “non-permitted” features a sufficient
approach to ensure fairness? (again, assume the models
are accurate, etc.)

10. Fairness implemented:
- Can we identify a such a feature set?
- How do we select these features? (legal framework, …)
- What if the non-permitted features are directly related to
permitted features?

11. Fair?
Decision to change a medical treatment – to the benefit of
the patient – based on race.

12. Fairness implemented:
- Can our permitted feature set be context independent?
Or does each problem need its own list?
- Societal vs. personal gains?

13. Discussion points:
- Can fairness be formalized?
- Is the current legal framework sufficient?
- Can we enforce / regulate fairness?
- …?