The concern about underlying discrimination hidden in machine learning (ML) models is increasing, as ML systems have been widely applied in more and more real-world scenarios and any discrimination hidden in them will directly affect human life. Many techniques have been developed to enhance fairness including commonly-used group fairness measures and several fairness-aware methods combining ensemble learning. However, existing fairness measures can only focus on one aspect -- either group or individual fairness, and the hard compatibility among them indicates a possibility of remaining biases even if one of them is satisfied. Moreover, existing mechanisms to boost fairness usually present empirical results to show validity, yet few of them discuss whether fairness can be boosted with certain theoretical guarantees. To address these issues, we propose a fairness quality measure named discriminative risk to reflect both individual and group fairness aspects. Furthermore, we investigate the properties of the proposed measure and propose first- and second-order oracle bounds to show that fairness can be boosted via ensemble combination with theoretical learning guarantees. The analysis is suitable for both binary and multi-class classification. A pruning method is also proposed to utilise our proposed measure and comprehensive experiments are conducted to evaluate the effectiveness of the proposed methods.
翻译:暂无翻译