On Fairness and Calibration
This work addresses fairness and calibration issues in machine learning models, revealing fundamental limitations that impact practitioners and researchers, though it is incremental as it extends existing results.
The paper investigates the conflict between achieving fairness in classification by minimizing error disparities across groups and maintaining calibrated probability estimates, showing that calibration is only compatible with a single error constraint and that such algorithms are no better than randomizing predictions, with empirical confirmation on datasets.
The machine learning community has become increasingly concerned with the potential for bias and discrimination in predictive models. This has motivated a growing line of work on what it means for a classification procedure to be "fair." In this paper, we investigate the tension between minimizing error disparity across different population groups while maintaining calibrated probability estimates. We show that calibration is compatible only with a single error constraint (i.e. equal false-negatives rates across groups), and show that any algorithm that satisfies this relaxation is no better than randomizing a percentage of predictions for an existing classifier. These unsettling findings, which extend and generalize existing results, are empirically confirmed on several datasets.