Bias Meets Bayes: A Bayesian Perspective on Improving Model Fairness
summary
Simulation uses a 90%/10% split with 450 majority and 50 minority samples and a data generator whose true risk rate is 36%. A vanilla logistic model (no group) yields false positive rate ~28% for the minority, downsampling the majority worsens it to ~41–42%, while a Bayesian logistic with an explicit group coefficient and a conservative prior implemented in PyMC shrinks the minority FPR to ~10%, keeps accuracy similar, improves calibration, and yields a group-parameter 95% CI spanning −0.11 to 0.19.
AI-written summary · about
chapters (11)
- 0:00Fairness and Bayesian stats
- 1:57Biased model predictions
- 4:26Fairness and bias defined
- 6:58Fairness metrics importance
- 7:41Bayesian framework for fairness
- 10:28Simulation setup for fairness
- 15:27Fairness through unawareness
- 18:34Data resampling impact on fairness
- 20:00Bayesian bias mitigation
- 24:51Advanced structural fairness models
- 26:16Q&A
related talks
01h 30mFairness Tales: How To Measure And Mitigate Unfair Bias in Machine Learning Models
46mThe Measure and Mismeasure of Fairness
30mBuilding fairer models for finance
39mBeyond Bias: Crafting Fairness in Pythonic Machine Learning Systems
28mMeasuring Model Fairness
50mThe Measure and Mismeasure of Fairness
38mTracking and Improving Information in the Service of Fairness
31mMeasuring Model Fairness