The Elephant in your Dataset Addressing Bias in Machine Learning
summary
Overview of measurable failures and fixes: explains that the Compass recidivism dataset yields similar overall accuracy but Black defendants receive false positives about twice as often, demonstrates using the published ProPublica data and Microsoft R|Widgets in a Jupyter notebook, and shows that simple blindness (dropping race) fails due to Simpson’s paradox. Shows pre-processing and balancing the target rates can equalize demographic parity but costs roughly a 5% drop in accuracy, arguing for multi-objective optimization plus governance.
AI-written summary · about
chapters (15)
- 0:00🚀 Build Stuff Conference – Get Your Tickets!
- 0:25Introduction: The Importance of Addressing Bias in AI
- 3:10The Wheel of Privilege and Intersectionality
- 6:15What is Bias? An Overview
- 10:20Types of Bias in Machine Learning Systems
- 13:00How Harmful Bias Affects AI Outcomes
- 17:00Real Examples: Historical Bias and Representation Bias
- 22:30Measurement Bias: A Deep Dive
- 27:00Case Study: Compass System and Recidivism Prediction
- 32:10Bias in Healthcare and Other Real-World Applications
- 34:00The Bias-Variance Tradeoff and Fairness
- 37:30Methods for Measuring Bias in AI
- 39:50Demo: Pre-processing and Balancing Data
- 43:10Addressing Bias Through Multi-objective Optimization
- 45:00The Socio-technical Approach to Mitigating Bias
related talks
01h 00mThe Elephant in your Dataset: Addressing Bias in Machine Learning
43mThe Elephant in your Dataset: Addressing Bias in Machine Learning
55mEthical Machine Learning
47mEthical Machine Learning
38mSocial Implications of Bias in Machine Learning
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....
23mRemoving Unfair Bias in the Machine Learning Pipeline
38mSocial Implications of Bias in Machine Learning