Time Series Forecasting and Beyond: Efficient Modeling, Structured Learning, and Symbolic Reasoning

摘要

Time series forecasting has become a central problem in modern machine learning, with applications spanning science, engineering, health, and finance. Yet in many domains, accurate prediction is only the starting point. This talk explores a broader view of time series learning, moving from forecasting toward discovery. I will first discuss our work on efficient time series forecasting, focusing on methods that improve scalability and practicality without sacrificing predictive performance. I will then turn to structured learning for high-dimensional temporal data, showing how modeling dependencies across variables can lead to more reliable and expressive forecasting systems. Finally, I will ask whether machine learning can move beyond forecasting to discover symbolic structure from time series data. I will present our recent work on benchmarking symbolic reasoning over time series and on hybrid LLM-based methods for scientific discovery. These directions point toward time series models that support both prediction and discovery.