Engineering Fair Machine Learning Pipelines

Abstract

Data splits and data preparation during fairness mitigation are known to influence the performance of output models. We propose including protected attributes in stratification when splitting a dataset. We also describe fairness patterns for assembling fair pipelines that include data preparation, estimators, and mitigators. This paper introduces an open-source Python library lale.lib.aif360 that offers sklearn compatible implementations of fair stratification and fairness patterns.