Predict Responsibly: Improving Fairness and Accuracy by Learning to Defer

In many machine learning applications, there are multiple decision-makers, both automated and human. The interaction between these agents often unaddressed in algorithmic development. In this work, we explore a simple of this interaction with a two-stage framework containing an automated and an external decision-maker. The model can choose to say "Pass", and the decision downstream, as explored in rejection learning. We extend this by proposing "learning to defer", which generalizes rejection learning considering the effect of other agents in the decision-making process. We a learning algorithm which accounts for potential biases held by decision-makers in a system. Experiments demonstrate that learning to can make systems not only more accurate but also less biased. Even when with inconsistent or biased users, we show that deferring models still improve the accuracy and/or fairness of the entire system.