Grant: Discovery Grants Program - Individual

With prediction and optimization models becoming more ingrained in decision making processes across multiple sectors (e.g., healthcare, criminal justice, finance), algorithmic advice serves as a valuable means to reduce biases in those processes. Current efforts to do so focus on developing algorithmic fairness notions, which are used to create algorithmically fair models that do not have "any prejudice or favoritism". However, in high-stakes settings, algorithms only guide human experts, who can bypass algorithmic advice to make decisions. This means that focusing only on fairness within algorithms alone is not enough to develop fair systems as it fails to address biases from human decision makers. My program's long-term goal is to create operations research and machine learning (ML) frameworks to improve fairness in high-stakes settings (e.g., healthcare) that involve human-algorithm decision making. My program's short-term objectives aim to mitigate the harms of biased humans in such processes by developing new assignment optimization frameworks, where human-algorithm decision makers (e.g., physicians), with varying predictive abilities, are assigned subjects (e.g., patients) to evaluate for a given prediction task. These assignments will be optimized to improve the overall predictive performance across each decision maker, while ensuring fairness (defined as similar prediction errors across groups and individuals that are being evaluated). The assignment optimization framework will be developed for settings where a human's adherence to algorithmic advice: (i) can be guided and (ii) cannot be guided and instead must be learned. The former requires modeling human-algorithm interactions into optimization models that optimize fairness, predictive performance, and wait times for the decision makers. The latter requires a new active learning and optimization framework for collecting real-world data to develop a prediction model that estimates how human-algorithm decision makers make decisions, which will then be used to make optimal assignments. Finally, to further improve a system's resiliency to biased decision makers, I will develop a new algorithmic fairness notion that accounts for biases from previous decisions using an optimization framework. This new fairness notion will be integrated into the assignment optimization frameworks to ensure such resiliency translates to practice. This program will advance and connect the study of algorithmic fairness and human-algorithm decision making, bridging the research-to-practice disconnect in improving fairness through algorithms. It also develops and presents a new paradigm to algorithmic fairness research, namely, assignment optimization to mitigate human biases in human-algorithm decision making. Together, this research program represents a Canadian led innovation in algorithmic fairness research and further establishes Canada as a research leader in operations research and ML.