Modeling Temporal and Structural Information in Time Series Data

Time series data is a sequence of data with temporal information at each position in the sequence. Such data widely exists in various disciplines. In computer science, different areas such as computational biology, signal processing, anomaly detection, and user behavior modeling benefit significantly from time series data. When modeling and analyzing time series data, there are two essential aspects embedded in time series data. The first one is structural information. Structural information contains the relationships and dependencies that inherently exist in time series data. The second indispensable aspect is temporal information. Temporal information is the key to distinguish time series data from other sequence data such as sentences (sequences of words). This thesis proposes novel approaches for modeling structural and temporal information to improve performance on various machine learning tasks. It demonstrates that the same methodologies can be used for diverse machine learning tasks, including activity recognition, dynamic network prediction, hypothesis testing, and recommendation. For activity recognition, I propose a novel adversarial prediction approach to model structured outputs, which outperforms the state-of-the-art approaches. I also design adversarial structural prediction approach that provides robust guarantees and superior performance for dynamic network prediction on real-world network prediction datasets. Additionally, I demonstrate that new temporal features are capable of capturing favorable information for the dynamic network prediction task. Another proposed approach in this thesis is temporal filtering which is introduced to advance the learning tasks of hypothesis testing and recommendation.