Fair for groups, unfair for individuals: Closing the gap in AI fairness.
- WAI CONTENT TEAM
- 2 days ago
- 2 min read
Original paper: Stairway to Fairness: Connecting Group and Individual Fairness
Authors: Theresia Veronika Rampisela, Maria Maistro, Tuukka Ruotsalo, Falk Scholer, & Christina Lioma
Read the paper: https://dl.acm.org/doi/10.1145/3705328.3748031
About the researcher
Theresia Veronika Rampisela, Postdoctoral Researcher, University of Copenhagen, Denmark. Find out more here: https://www.linkedin.com/in/theresia-rampisela/
What problem does this paper address, and why does it matter?
Fairness for recommender system users can be evaluated for groups (e.g., based on users' demographic attributes) and for individuals (e.g., without considering their demographic attributes). However, the relationship between them is not well-understood, as prior work on both types has used different evaluation measures or evaluation objectives for each fairness type, thereby not allowing for a proper comparison of the two. As a result, it is currently not known how improving one type of fairness may affect the other.
What did this research discover/create?
For a more apple-to-apple comparison between individual and group fairness, we first identify metrics that can be used to evaluate both of fairness types and compute them in our experiments. Our key findings show that group fairness measures often hide unfairness within (intersectional) groups and between individuals.
How could this research impact real-world applications?
This research could impact real-world applications by showing the importance of evaluating for multiple fairness definitions, to ensure that the applications are fair not only towards various groups, but also for all individuals. By encouraging researchers, practitioners, and policymakers to evaluate individual fairness on top of group fairness, our work pushes towards efforts and regulations on mitigating unfairness across individuals, such that all users can obtain a similar level of utility.
Who should care about this work?
Fairness researchers, recommender system practitioners, and policymakers setting fairness criteria for high-risk AI applications.
What is noteworthy about this research?
The insight is unique as most fairness research involving recommender systems focus exclusively on group fairness, less on individual fairness, and very rarely on both fairness types, which is what our work does. In this way, we can have a comparable way of measuring group fairness and individual fairness.
What's the ONE key takeaway you want people to remember?
Evaluating fairness beyond the between-group level is important, as AI applications (in this case, recommender systems) can be relatively fair for groups and at the same time much more unfair for individuals.
Funding/sponsorship information for this research
This work is supported by the Algorithms, Data, and Democracy project (ADD-project), funded by the Villum Foundation and Velux Foundation. This work is also supported in part by the Australian Research Council (DP190101113).
Read More
Access the full paper: https://dl.acm.org/doi/10.1145/3705328.3748031
Have research to share or recommend? https://forms.gle/kaTJEd3gfJV8vtDx8
Drive Mad offers a colorful and simple visual style that keeps attention on the vehicle, obstacles, and track ahead. In comparison, Mad Drive can deliver a similar arcade atmosphere while giving players another way to experience challenging car-based gameplay.