Empirical analysis of fairness-aware data segmentation

Abstract

Fairness in machine learning is a research area that is recently established, for mitigating bias of unfair models that treat unprivileged people unfavorably based on protected attributes. We want to take an approach for mitigating such bias based on the idea of data segmentation, that is, dividing data into segments where people should be treated similarly. Such an approach should be useful in the sense that the mitigation process itself is explainable for cases that similar people should be treated similarly. Although research on such cases exists, the question of effectiveness of data segmentation itself, however, remains to be answered. In this paper, we answer this question by empirically analyzing the experimental results of data segmentation by using two datasets, i.e., the UCI Adult dataset and the Kaggle ‘Give me some credit’ (gmsc) dataset. We empirically show that (1) fairness can be controllable during training models by the way of dividing data into segments; more specifically, by selecting the attributes and setting the number of segments for adjusting statistics such as statistical parity of the segments and mutual information between the attributes, etc. (2) the effects of data segmentation is dependent on classifiers, and (3) there exist weak trade-offs between fairness and accuracy with regard to data segmentation.