Machine Learning Model Fairness and Interpretability with AWS
summary
Defines concrete sources of bias (sampling, measurement, missing features, human selection), shows a 2,000-example fraud dataset with only five fraud cases to justify rebalancing, and demonstrates SHAP/Shapley values for per-instance and global explanations. Walks through using Amazon SageMaker Clarify to detect bias, produce local and global explanations, and monitor data drift and model drift over time, and describes integrating Clarify into an MLOps pipeline for continuous auditing and retraining.
AI-written summary · about
chapters (9)
- 0:00ML fairness intro
- 3:25Bias in data and ML
- 9:20Types of ML bias
- 14:18Detecting and fixing bias
- 20:03Fairness accountability
- 26:27Explaining ML decisions
- 31:51SageMaker Clarify for bias
- 35:41Clarify in ML Ops
- 37:28Summary on ML fairness
related talks
44mImprove explainability of ML models to meet regulatory requirements
50mEthical Machine Learning
55mEthical Machine Learning
24mUnderstand ML model predictions & biases with Amazon SageMaker Clarify
18mUnderstand your ML models better with greater visibility and transparency
33mBuild trustworthy ML detection and monitor bias
36mMachine Learning Fairness: Lessons Learned
54mEthical Machine Learning