Accurate Fairness: Improving Individual Fairness without Trading Accuracy

Accuracy and individual fairness are both crucial for trustworthy machine learning, but these two aspects often incompatible with each other so that enhancing one aspect may sacrifice the inevitably side effects of true bias or false fairness. We propose in this paper a new criterion, accurate fairness, to align accuracy. Informally, it requires treatments an individual's similar counterparts conform uniform target, i.e., ground truth individual. prove also implies typical group criteria over union sub-populations. then present Siamese in-processing approach minimize accuracy losses learning model under constraints. To best our knowledge, is first time adapted mitigation. confusion matrix-based metrics, fair-precision, fair-recall, fair-F1 score, quantify trade-off between Comparative case studies popular datasets show can achieve on average 1.02%-8.78% higher (in terms through awareness) 8.38%-13.69% accuracy, as well 10.09%-20.57% fair rate, 5.43%-10.01% than state-of-the-art mitigation techniques. This demonstrates indeed improve without trading Finally, criterion applied mitigate possible service discrimination real Ctrip dataset, by fairly serving 112.33% more customers (specifically, 81.29% accurately way) baseline models.