Towards model-based bias mitigation in machine learning

Models produced by machine learning are not guaranteed to be free from bias, particularly when trained and tested with data produced in discriminatory environments. The bias can be unethical, mainly when the data contains sensitive attributes, such as sex, race, age, etc. Some approaches have contributed to mitigating such biases by providing bias metrics and mitigation algorithms. The challenge is users have to implement their code in general/statistical programming languages, which can be demanding for users with little programming and fairness in machine learning experience. We present FairML, a model-based approach to facilitate bias measurement and mitigation with reduced software development effort. Our evaluation shows that FairML requires fewer lines of code to produce comparable measurement values to the ones produced by the baseline code.