Fairness Tales: How To Measure And Mitigate Unfair Bias in Machine Learning Models
summary
Workshop generates a synthetic 400 CV dataset (OpenAI) and a biased recruiter simulator that produces a 41% callback baseline and a male callback rate ~70% vs female ~10%. A LightGBM text classifier learns that injected word “woman” and reproduces the bias; mitigation uses Fairlearn threshold optimizer and reductions, plus audits with Aequitas (and mention of IBM Fairness 360) to push equalized odds from ~78 to ~55 then ~33 while accuracy falls from ~81% to ~73%.
AI-written summary · about
chapters (11)
- 0:00Fairness in AI hiring
- 10:47AI fairness history and law
- 21:28Workshop setup and data
- 33:52Generating biased CVs
- 44:59Biased text classifier
- 53:41Fairness metrics and challenges
- 1:05:02Aequitas fairness audit
- 1:08:50Legal disparate impact rules
- 1:09:33Mitigating bias with Fairlearn
- 1:21:29Fairness-accuracy trade-off
- 1:24:40AI fairness and regulations
related talks
01h 31mHow To Measure And Mitigate Unfair Bias in Machine Learning Models
23mRemoving Unfair Bias in the Machine Learning Pipeline
39mBeyond Bias: Crafting Fairness in Pythonic Machine Learning Systems
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
42mFairness in Machine Learning: From Theory to Practice
47mEthical Machine Learning
51mDigital Discrimination: Cognitive Bias in Machine Learning
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....