What Can Program Analysis Say About Data Bias?
summary
Proposes a method that uses abstract interpretation to symbolically represent all datasets within a bias model (e.g., up to n label flips) and lift core operations so a decision-tree learner is trained over that set, producing intervals of outcomes instead of single labels. Implements the analyzer, certifies robustness on fairness benchmarks (e.g., 73% certified at 1% missing-data bias for Adult Income), finds some subgroups lose certification quickly (black groups drop near 0.4%), and contrasts a bagging-based probabilistic guarantees approach that scales but loses interpretability.
AI-written summary · about
chapters (10)
- 0:00Start
- 8:10Models of data bias
- 16:28Abstraction insight
- 18:26Example: Abstract dataset
- 21:16Example: Counting
- 26:13Example: Ratio/probabiltiy operator
- 27:32Learning a decision tree
- 30:03Abstract tree learning
- 41:28Evaluation A: Certification accuracy
- 45:46Evaluation B: Certification disparity
related talks
47mFight AI Bias With… Bias
30mBias, Explainability, and Accountability: The Data Scientist's Burden
31mFlagging *and* Fixing Bias in ML
46mA look at the methods to detect and try to remove bias in ML models
42mUnderstanding the nature of Bias in Natural & AI Systems
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....
55mEthical Machine Learning
26mData Bias: The Danger Lies in the Details