Facilitating Human-Machine Teaming in Machine Learning Problem Formulation

The first step in any applied machine learning (ML) task is formulating what needs to be predicted and how to define it, considering the available data. It often takes significant trial-and-errors for practitioners to identify a problem formulation that is both relevant to the task and realistic given the available data. Despite its importance, this process is not well supported by existing tools. This is a task where humans and machines bring complementary expertise, thus potentially amenable to human-machine teaming. My research was focused on answering a core research question: How can we facilitate problem formulation through human-machine teaming and what's the impact of human-machine teaming. We first conducted a study to investigate the impact of two human-machine teaming strategies: 1) Performance First: machines leading the process by recommending problem formulations based on predictive performance, and 2) Relevance First: humans leading the process by selecting relevant proxies. Then, we conducted a second study to understand the impact of different machine recommendation strategies: 1) recommend based on relevance, 2) recommend based on predictive performance, and 3) recommend Pareto optimal problem formulations by considering it as a multi-objective optimization task. Across two studies, introducing automation to facilitate rapid prototyping increases the number of candidates that users can examine. However, we also observed a consistent bias towards quantitative metrics of model quality, risking misalignment with application goals. We show that systems should explicitly support the evaluation of user objectives and facilitate multi-criteria tradeoffs during the problem formulation process. Together, these insights help provide a concrete step toward interactive problem formulation tools that not only accelerate iteration but also support multi-objective exploration.