Fairness in Machine Learning: From Theory to Practice
summary
Talk argues that perfect fairness is impossible but actionable steps reduce harm: measure subgroup performance using sensitivity and specificity, pick thresholds to trade off false positives vs false negatives, and apply pre/in/post‑processing to remove or constrain bias. Cites real domains—health insurance, credit scoring, political models—where skewed samples and proxies (zip code, education) produce disparate outcomes. Recommends adversarial testing, documentation, participatory design, reproducible tooling like Fairlearn and tensorflow indicators, and using post-processing thresholds for remediation.
AI-written summary · about
chapters (5)
- 0:00Start
- 18:20Specificity VS Sensitivity
- 22:26Preventing Unfairness in Machine Learning Models
- 26:31How do we Apply an Open Source Framework to ML Practices?
- 33:16Dangers of Policy Pitfalls
related talks
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
36mMachine Learning Fairness: Lessons Learned
47mEthical Machine Learning
51mDigital Discrimination: Cognitive Bias in Machine Learning
23mRemoving Unfair Bias in the Machine Learning Pipeline
42mFairness and transparency in machine learning: Tools and techniques
31mFlagging *and* Fixing Bias in ML
35mBuilding Inclusive Machine Learning Models