Beyond evaluation: Improving fairness with Model Remediation
summary
Demo applies MinDiff model built around Keras to a Civil Comments dataset and claims measurable reduction in false positive rates for religious identity slices. It selects minority and majority negative examples, adds a MinDiff loss with a tunable weight, trains without enlarging the dataset, and reports narrower confidence intervals and lower FPRs across all shown subgroups while noting results can vary and parameters must be rechecked.
AI-written summary · about
chapters (4)
related talks
47mEthical Machine Learning
46mA look at the methods to detect and try to remove bias in ML models
28mMeasuring Model Fairness
42mFairness in Machine Learning: From Theory to Practice
31mMeasuring Model Fairness
39mBeyond Bias: Crafting Fairness in Pythonic Machine Learning Systems
48mResponsible AI
31mBias Meets Bayes: A Bayesian Perspective on Improving Model Fairness