Pipelines for Fairness: A Convexing Usecase
summary
Pipeline approach presents concrete fixes and measured tradeoffs: a modified Gram-Schmidt transformer removes linear dependence on protected columns inside a scikit‑learn pipeline, a convexified constraint implemented via a convex modeling language solves for parameters under an equal‑opportunity covariance bound, and a meta‑estimator postprocesses thresholds per subgroup. Experiments use the COMPAS-style dataset with about 70% accuracy and raise equal opportunity from around 0.82 gender 0.78 race while accuracy drops only a percent or so.
AI-written summary · about
chapters (8)
- 0:00Fair ML pipelines intro
- 1:08Why fairness matters in machine learning
- 3:57Measuring fairness
- 10:24Remove sensitive info
- 13:19Optimize with fairness constraints
- 15:33Solve convex fairness
- 18:49Post-process fairness
- 22:48Ethical limits
related talks
41mFairness for Scikit-Learn Pipelines with Lale
23mRemoving Unfair Bias in the Machine Learning Pipeline
42mFairness and transparency in machine learning: Tools and techniques
28mLife is not fair, but your ML pipeline can be!
01h 31mHow To Measure And Mitigate Unfair Bias in Machine Learning Models
39mBeyond Bias: Crafting Fairness in Pythonic Machine Learning Systems
46mA look at the methods to detect and try to remove bias in ML models
34mInterpretability and Algorithmic Fairness