Learning When to Ask Questions about the Current Context

Mobile phones, being ubiquitous and always accessible, provide an opportunity for symbiotic interaction between humans and machines, enabling the understanding of human behavior at any time and from any location through sensor data. While sensor data offers an objective view of reality, it fails to capture the subjective motivations behind an individual's behavior; to learn the subjective motivations, the general way is to ask context questions to the user. However, a key problem was raised because the answers from people are always not of good quality. This thesis aims to solve the problem of inadequate human input quality, typically marked by a high number of wrong answers, for three personas: researchers, participants, and the platform. For researchers, the thesis demonstrates how users' reaction times before starting to respond significantly affect answer quality, and it designs a comprehensive methodology for enabling participants to provide high-quality input data. For participants, it identifies optimal moments or situational contexts when they are most willing to provide high-quality input data. For the platform, it designs a novel system to automatically collect high-quality human input data, comprising a smartphone application for sensor data and user feedback collection, a calendar for human-machine collaboration configuration, a dashboard for qualitative and quantitative experiment progress feedback, user-centric notifications, and a machine-learning module (pbgc-Forest algorithm) to facilitate human-machine interactions. This thesis paves the way for new opportunities in mobile user interfaces and interaction technologies to enhance user experiences and improve the quality and reliability of data collected from participants, ensuring more accurate and meaningful research outcomes.