Abstract
In spatiotemporal systems, large amounts of time series data—such as urban mobility and climate variables—are readily available for implementing downstream machine learning tasks and supporting decision-making. Since these data are characterized by spatiotemporal dimensions and reflect underlying system patterns, it is essential to reformulate time series problems using machine learning approaches. In this talk, we address two key scientific challenges: 1) Imputing missing values from partially observed time series data; 2) Quantifying time series periodicity in spatially- and time-varying systems. To tackle the first problem, we introduce a time series convolution framework that efficiently imputes missing data while maintaining scalability through the Fast Fourier Transform (FFT). For the second problem, we propose a mathematical framework based on sparse autoregression and mixed-integer programming to quantify periodicity in multidimensional mobility data and climate variables. Our models would lay an insightful foundation for understanding complicated spatiotemporal data in real-world systems.
For more info, please follow this link.