A look at the methods to detect and try to remove bias in ML models
summary
Survey of fairness methods compares pre processing, in processing and post processing approaches and shows concrete results on COMPAS, UCI Adult and a credit dataset: using reweighing, optimized pre-processing and the disparate impact remover moves statistical parity difference closer to zero and disparate impact toward 1 while accuracy drops only about 1–3 percentage points (e.g., 77% → 75–76%). Demonstrations use IBM AIF360 and adversarial or regularizer-based in‑processing and equalized‑odds postprocessing; balancing training data often fixes vision models by collecting more samples.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 0:41Data Privacy
- 2:39Impact on people
- 6:35Pricing Sentence
- 9:32Favorable Label
- 11:51Bias
- 17:32Disparate impacts
- 18:54Proportional parity
- 22:56Three steps
- 24:12Three paths
- 25:45Preprocessing
- 27:34Processing
- 28:53Postprocessing
- 30:06Experiments
- 32:24Results
- 34:24Comparing results
- 36:48Tradeoff in accuracy
- 40:20Unstructured data
related talks
23mRemoving Unfair Bias in the Machine Learning Pipeline
51mDigital Discrimination: Cognitive Bias in Machine Learning
31mFlagging *and* Fixing Bias in ML
42mFairness in Machine Learning: From Theory to Practice
28mMeasuring Model Fairness
39mBeyond Bias: Crafting Fairness in Pythonic Machine Learning Systems
21mFairness and bias in AI
47mFight AI Bias With… Bias