Understanding the Effects of Mitigation on De-identified Data

Abstract

Machine learning algorithms can play a significant role in peoples lives. The data used in these algorithms often contains sensitive information and can have inherent biases. We investigate the effects of the interaction between mechanisms designed to preserve privacy and to mitigate biases. The mechanisms employed for this investigation were k-anonymity for de-identification and re-weighting for bias mitigation. The experiments were threefold. First, we investigated the effects of mitigation on de-identified data. Second, we measured the effects of three data parameters: class imbalance ratio, privileged positive outcome ratio and unprivileged positive outcome ratio. Third, we assessed the utility of the mitigation mechanism in a healthcare context. Using real-world data, we tested the effects of different levels of de-identification. We primarily utilised three measures to indicate the procedures’ effects. First, for accuracy, we analysed simple accuracy and balanced accuracy rate. Second, we measured the number of positive outcomes for the privileged and unprivileged class and the disparate impact for fairness. Third, for utility, we measured the recall rate as well as a novel metric; recall ratio. We display these two metrics based on the classification threshold to indicate the trade-off between achieving high true positives while limiting overall positive outcomes. This trade-off is analogous to a medical testing scenario where the objective is to have high accuracy of true positives and minimise cost given overall percent positives.