Computational Method for Understanding Complex Human Routine Behaviors

The ability to collect and store large amounts of human behavior traces data collected from various sensors on people’ personal, mobile, and wearable devices, as well as from smart environments, offers a new source of data to study human behavior at scale. However, existing Human-Computer Interaction (HCI) behavior sensemaking methodologies do not lend themselves to studying behaviors from such large multivariate, heterogeneous, and
unlabeled datasets. On the other hand, computational modeling has been used to successfully explore and understand complex systems in other fields (e.g., climate change modeling). Inspired by such prior work, we treat behaviors stored in large behavior logs as
a complex system that we capture in a computational model of human behavior. In this work, we focus on behaviors in the domain of human routines that people enact as
sequences of actions they perform in specific situations, which we call behavior instances. Computational models then allow us to explore different kinds of behaviors by
manipulating model variables and simulating and detecting different kinds of behaviors (otherwise known as “asking what-if questions”). In this thesis, we propose a probabilistic
computational model of human routine behaviors, that can describe, reason about, and act in response to people’s behaviors. We ground our model in a holistic definition of human routines to constrain the patterns it extracts from the data to those that match routine behaviors. We train the model by estimating the likelihood that people will perform certain actions in different situations in a way that matches their demonstrated preference for those
actions and situations in behavior logs. We leverage this computational model of routines to create various tools to aid stakeholders, such as domain experts and end users, in
exploring, making sense of, and generating new insights about human behavior stored in large behavior logs in a principled way.