Time series analysis is essential for a wide range of fields, ranging from business intelligence to healthcare. Among various time series data, two representative types of time series are temporally-ordered wearable sensory data (e.g., heart rate) and human behavioral data (e.g., purchase behavior). The key challenge in time series data is to comprehensively capture the underlying temporal pattern from sequential historical observations. To achieve this goal, this dissertation aims to develop novel deep learning frameworks, which explore the dynamic temporal patterns for different applications.
Works in this dissertation explore a variety of applications in analyzing different types of time series data. In particular, we first develop novel deep learning models for representation learning on wearable sensory time series data, by automatically mapping any variable-length series into the low-dimensional latent space which effectively preserves the most informative temporal contextual signals and relevant inherent properties. Beyond the general time series representation learning task, we also investigate the time series imputation problem with the exploration of cross-series dependencies. Additionally, investigating human behavioral data is another key dimension which is in pressing need for time series analysis. To address this challenge, we propose three new learning architectures which capture temporal dynamics from different perspectives, including predefined resolution-aware forecasting, automated resolution-aware forecasting and customized forecasting.