Machine learning to complement human decision making

Résumé

In many settings, machine learning models are used in conjunction with a human decision maker. For example, consider a model used to make diagnoses in healthcare. Far from having the model operate autonomously, human clinicians may provide a second opinion on hard cases, use the predictions to decide on followup actions, and so on. This raises a challenge: how can we design machine learning models which best complement the strengths, weaknesses, and decision-making process of humans? I will start by discussing the triage setting, where the model must decide whether to query a (costly) human expert for help with a given example. We find that explicitly modeling the human's response, and adapting the model to reflect their strengths and weaknesses, yields better team performance than a model trained for accuracy in isolation. Then, I will turn to uncertainty quantification, where the model outputs a set of possible labels for use by a human instead of just a point prediction. Standard methods for producing prediction sets focus on ensuring coverage of the true label, agnostic to how useful a set is for a downstream decision maker. We propose a method to optimize prediction sets relative to a utility function that models a human's decision-making process, for example preferring sets that imply similar followup actions for clinician. Empirically, we find that our method produces prediction sets that are more cohesive and clinically interpretable while retaining the accuracy guarantees as before.