Generative Adversarial Networks for Mitigating Biases in Machine Learning Systems

In this paper, we propose a new framework for mitigating biases in machine\nlearning systems. The problem of the existing mitigation approaches is that\nthey are model-oriented in the sense that they focus on tuning the training\nalgorithms to produce fair results, while overlooking the fact that the\ntraining data can itself be the main reason for biased outcomes. Technically\nspeaking, two essential limitations can be found in such model-based\napproaches: 1) the mitigation cannot be achieved without degrading the accuracy\nof the machine learning models, and 2) when the data used for training are\nlargely biased, the training time automatically increases so as to find\nsuitable learning parameters that help produce fair results. To address these\nshortcomings, we propose in this work a new framework that can largely mitigate\nthe biases and discriminations in machine learning systems while at the same\ntime enhancing the prediction accuracy of these systems. The proposed framework\nis based on conditional Generative Adversarial Networks (cGANs), which are used\nto generate new synthetic fair data with selective properties from the original\ndata. We also propose a framework for analyzing data biases, which is important\nfor understanding the amount and type of data that need to be synthetically\nsampled and labeled for each population group. Experimental results show that\nthe proposed solution can efficiently mitigate different types of biases, while\nat the same time enhancing the prediction accuracy of the underlying machine\nlearning model.\n