Designing for Effective End-User Interaction with Machine Learning

Abstract

End-user interactive machine learning is a promising tool for enhancing human capabilities with data. Recent work has shown that we can create specific applications that employ end-user interactive machine learning. However, we still lack a generalized understanding of how to design effective end-user interaction with machine learning. This dissertation advances our understanding of this problem by demonstrating effective end-user interaction with machine learning in a variety of new situations and by characterizing the design factors affecting the end-user interactive machine learning process itself. Specifically, this dissertation presents (1) new interaction techniques for end-user creation of image classifiers in an existing end-user interactive machine learning system called CueFlik, (2) a novel system called ReGroup that employs end-user interactive machine learning for the purpose of access control in social networks, (3) a novel system called CueT that supports end-user driven machine learning for computer network alarm triage, and (4) a novel design space characterizing the goals and constraints impacting the end-user interactive machine learning process itself. Together, these contributions can move us beyond ad-hoc designs for specific applications and provide a foundation for future researchers and developers of end-user interactive machine learning systems.