Fairness for Scikit-Learn Pipelines with Lale
summary
Experiments use the credit G dataset (1,000 rows, ~20 features) with protected attributes age and personal status and report concrete numbers: dummy classifier hits perfect statistical parity but low balanced accuracy; mitigators—disparate impact remover, reweighing, calibrated equalized odds, and reject option classification—shift predictions toward fairness but none perfectly trade off accuracy and fairness and error bars are large; stratified splits reduce instability; Lale integrates scikit‑learn and aif360 components and AutoML explores Pareto frontiers over encoders, mitigators, estimators, and hyperparameters.
AI-written summary · about
chapters (16)
- 0:00Fairness in ML pipelines
- 1:08Algorithmic fairness matters
- 3:12Protected attributes in data
- 5:21Fairness and bias worldviews
- 8:51Feature importance and bias
- 12:22Fairness metrics essential
- 15:11Five fairness metrics trade-offs
- 20:01Pipeline fairness vs accuracy
- 24:04Bias mitigators for models
- 27:13Lale mitigators in pipelines
- 29:21Mitigator performance comparison
- 31:07Categorical encoders impact
- 32:11Extending pipelines with ensembling
- 34:19AutoML for fairness
- 36:03Conclusion and tools
- 37:07Q&A
related talks
24mPipelines for Fairness: A Convexing Usecase
39mBeyond Bias: Crafting Fairness in Pythonic Machine Learning Systems
42mFairness and transparency in machine learning: Tools and techniques
23mRemoving Unfair Bias in the Machine Learning Pipeline
28mLife is not fair, but your ML pipeline can be!
01h 30mFairness Tales: How To Measure And Mitigate Unfair Bias in Machine Learning Models
01h 31mHow To Measure And Mitigate Unfair Bias in Machine Learning Models
42mFairness in Machine Learning: From Theory to Practice