Abstract
Increasing utilization of machine learning based decision support systemsemphasizes the need for resulting predictions to be both accurate and fair toall stakeholders. In this work we present a novel approach to increase a NeuralNetwork model's fairness during training. We introduce a family of fairnessenhancing regularization components that we use in conjunction with thetraditional binary-cross-entropy based accuracy loss. These loss functions arebased on Bias Parity Score (BPS), a score that helps quantify bias in themodels with a single number. In the current work we investigate the behaviorand effect of these regularization components on bias. We deploy them in thecontext of a recidivism prediction task as well as on a census-based adultincome dataset. The results demonstrate that with a good choice of fairnessloss function we can reduce the trained model's bias without deterioratingaccuracy even in unbalanced dataset.