UCLA

In this era of big data, massive data are generated from heterogeneous resources every day, which provides an unprecedented opportunity for deepening our understanding of complex human behaviors. Modeling human behaviors requires robust computational methods that can not only capture semantics and useful insights from sparse and heterogeneous data, but also unravel sophisticated human behaviors at different scales. In addition, the enormous data velocity and the unparalleled scale of deep models also pose significant challenges to efficiency.

In this dissertation, we demonstrate a collection of research results that systematically improve the ecosystem of human behavior modeling based on representation learning. For heterogeneous data in various settings, we present practical representation learning methods to effectively and efficiently capture their semantics. Moreover, these representation learning methods can actually fill a niche to comfortably model different behaviors with atomic, compositional, and explainable operations, thereby modeling human behaviors at different scales.As a result, our proposed approaches not only address various real-world challenges in diverse domains, but also present the potentials to adopt valuable domain knowledge into machine learning.