Modeling Human Behavior with Passive Sensing: Modeling Paradigms, Limited Data and Spatio-Temporal Patterns

Passively sensed human behavior modeling has been impactful in areas of activity recognition,
mental health, and well-being. However, modeling human behavior can be challenging due to the
problems of limited data, heterogeneity and the choice between modeling paradigms. Limited data
is a consequence of constraints in acquiring ground truth. While, heterogeneity in behavior patterns
can exist both within and between individuals. Moreover, the choice between universal and personalized paradigms for human behavior modeling can also be a critical decision for researchers. Yet, machine learning studies either explore a universal approach to modeling or a personalized paradigm. However, the same data is mostly not investigated from the perspective of both the approaches which could provide a better understanding of their effectiveness. In addition, when data is limited the models tend to overfit. Strategies that are often deployed in the current landscape anticipate datasets with over thousands of labeled data points. However, in scenarios where data labels are in the hundreds, strategies such as deep learning approaches, data augmentation, or regularization prove to be heavily constrained. Furthermore, majority of studies approach human behavior modeling by applying machine learning models focused on meta-level classifications. This type classification modeling focuses on the overarching levels of human behavior; for example, in depression, there are five levels: minimal, mild, moderate, moderately severe, and severe. Rarely are there attempts to uncover heterogeneous patterns within the levels of depression that exists in individuals or between them. This dissertation aims to compare and evaluate universal and personalized models by applying both the approaches to a mental health dataset pertaining to adolescents. Through the study I demonstrate that in human behavior modeling, personalization models perform better and also show the features that contribute towards the predictions. The insights from the study serve as the basis for a solution I develop for limited data conditions that are often experienced by researchers. I design and validate a multimodal machine learning ranking framework that ranks predictions in scenarios where datasets are limited to hundreds of observations and deep learning techniques perform poorly. Realizing the limited exploration of heterogeneity in passively sensed human behavior modeling, I propose a novel image representation of sensor data in this dissertation. The aim is to enable the identification of heterogeneous patterns that exist within and between individuals. Computational geometric algorithms inspire this image representation of multimodal time series behavioral sensor data. I validate the image representation through experimentation and substantiate the innovation as a building block for impactful applications in behavior sensor signal profiling and signal variation unconstrained by labels.