Abstract:
|
Despite recent advances in algorithmic fairness, methodologies for achieving fairness with generalized linear models (GLMs) have hardly been explored in general, despite their wide usage in practice. In this study, we consider a widely used fairness criterion equalized odds for GLMs. We prove that in the case of GLMs, both criteria can be achieved by equalizing the distribution of the linear components of the GLM. We propose to use maximum mean discrepancy to equalize the linear components' distributions. We also derive theoretical properties for the resulting fair GLM estimator. To empirically demonstrate the efficacy of the proposed fair GLM, we compare it with other well-known fairness-aware methods on extensive benchmark datasets for binary classification and regression. In addition, we demonstrate that the fair GLM can generate fair predictions for a wide range of response variables, other than binary and continuous outcomes.
|