I work on machine learning for large-scale time series and its applications. I am generally interested in sequential decision making, optimization and spatiotemporal modeling. My work has been applied to learning dynamical systems, intelligent transportation, and climate informatics. For more details, see my curriculum vitae and research statement.
Long-term forecasting in environments with nonlinear dynamics is highly challenging, since there exist long-term temporal dependencies, higher-order correlations and sensitivity to error propagation. We present a novel family of neural sequence models, that learn the nonlinear dynamics directly using higher order moments and high-order state transition functions.
Large-scale spatiotemporal data such as climate measurements, location-based social networks have complex structures. In this project, we develop a scalable framework based on low-rank tensor learning to analyze them. We propose a series of algorithms ranging from offline, online to memory efficient learning to solve the problem. We show that our framework can produce accurate predictions in a scalable way and discover interesting patterns in spatiotemporal data.
Given a series of road network snapshots, we propose a latent space model to capture its topological and temporal properties. We present an incremental online algorithm which sequentially and adaptively learns the spatiotemporal patterns. Our framework enables real-time traffic prediction on a large volumne of Los Angeles traffic sensor data.