Strategies for Selecting and Adapting Machine Learning Systems to Support Different Types of Experts

Abstract

Machine learning prediction and explanation systems offer the ability to learn meaningful representations and patterns in otherwise messy or complex data. These representations can then be used to predict outcomes, to improve human understanding, or to complete tasks more efficiently. As these systems become more prevalent, they are used by a wider variety of people for a wider variety of tasks, meaning more adaptation is required to ensure these systems can be utilized accurately and efficiently. Additionally, these systems are complex and are therefore easy to misuse and misinterpret, particularly when applied to new contexts by individuals without a deep background in the underlying algorithms used by the learning system. My dissertation explores strategies for selecting and adapting machine learning prediction and explanation systems to help support users of varying sets of expertise in utilizing these systems. These strategies are: aligning systems with user goals, retaining nuance in explanations, and imparting an appropriate level of trust of system outputs. I have explored these strategies through four projects: Problems and an Alternative to Single Explanation Aggregations, Predicting Blood Glucose Test Accuracy in ICU Patients, Predicting and Explaining an Imminent Dementia Diagnosis with Limited Data, and Flexible System for Efficient Goal-Directed Self-Tracking Analysis. Finally, I discuss how these strategies and lessons learned from the highly contextualized projects compare and contrast to guidelines set forth by existing frameworks to guide human-AI interaction and system design.