The 11th Mining and Learning from Time Series Workshop: From Classical Methods to LLMs

Conformal Prediction for Time Series: Beyond Exchangeability and Towards Multivariate Dependence

Conformal prediction (CP) provides a powerful, distribution-free framework for constructing prediction intervals with finite-sample coverage guarantees. However, most classical CP methods rely on the assumption of data exchangeability, which is typically violated in time series due to temporal dependence. This has motivated a growing body of work aimed at extending CP to the time-series setting. Another computational challenge comes from handling multi-dimensional time-series. In this talk, I will review recent advances in conformal prediction for time series, including our own contributions to a general framework for constructing distribution-free prediction intervals for time series that wrap around a given black-box algorithm. The new approach relaxes the exchangeability assumption and allows for general forms of temporal and multi-variate dependence across dimensions. Theoretically, we establish asymptotic marginal and conditional coverage guarantees, which comes with certain width optimality in some cases. Methodologically, we propose computationally efficient procedures based on ensemble predictors that are closely related to standard CP, yet tailored for time series.

Biography
Yao Xie is the Coca-Cola Foundation Chair and Professor in the H. Milton Stewart School of Industrial and Systems Engineering (ISyE) at the Georgia Institute of Technology. She also serves as Associate Director of the Machine Learning Center and holds an adjunct faculty appointment in the School of Electrical and Computer Engineering (ECE). Her research focuses on the intersection of statistics, optimization, and machine learning, with an emphasis on developing computationally efficient and statistically rigorous algorithms for solving engineering challenges drawn from real-world applications. Her specific areas of interest include sequential data analysis, change-point detection, high-dimensional data analysis, spatio-temporal modeling, and optimization for robust learning and inference. Her work has contributed to societal and policy advancements, including data-driven policing initiatives. Her research has been featured in outlets such as Inside Higher Ed and Georgia Tech's "In Pursuit of Better Data Forecasting" and "Driving Data Science Innovation as Coca-Cola Foundation Chair." Yao Xie holds an Erdős number of 3. In recognition of her research and impact, she received the INFORMS Donald P. Gaver, Jr. Early Career Award for Excellence in Operations Research in 2022. This prestigious award, given to a single recipient annually across INFORMS, cited her "outstanding research contributions at the interface of operations research, statistics, machine learning, and optimization; for successfully applying her research talent to applications of societal importance; and for contributions to the education and mentoring of students at all levels." In 2024, she was further honored with the CWS Woodroofe Award, presented at the Joint Statistical Meeting (JSM) 2024.

Towards Deep Time Series Modeling via Data-Centric AI

Time series data, as a pervasive kind of data format, have played an important role in numerous real-world scenarios such as energy, transportation, healthcare, etc. Effective time series modeling can help with accurate forecasting, resource optimization, and risk management. Given the great importance, how can we effectively model the nature of the pervasive time series data? This talk will focus on deep learning for time series modeling from a novel data-centric AI perspective. This perspective originally categorizes data-centric time series modeling into “learning” and “manipulating” time series data distribution. In “learning” time series data, I will investigate how to extract the distribution regularity to expand dependencies for forecasting. In “manipulating” time series data, I will discuss how to overcome the distribution shift for advancing generalization. By incorporating deep learning-based time series analysis with data-centric AI techniques, this talk aims to achieve more effective and efficient time series modeling and subsequent decision-making.

Biography:
Dr. Yanjie Fu is an associate professor in the School of Computing and AI at the Arizona State University. He received his Ph.D. degree from the Rutgers University in 2016, the B.E. degree from the University of Science and Technology of China (USTC), and the M.E. degree from the Chinese Academy of Sciences (CAS). His teaching and research have been recognized by various awards, including: US NAE Frontiers of Engineering early career engineer (2023), US NSF CAREER (2021), 2024 Stanford Elsevier World’s Top 2% Scientists, Fulton Engineering Top 5% Teaching Recognition Award, University Reach the Star award. He is committed to data science education. His graduated Ph.D. students have joined academia as tenure-track faculty members. His recent focuses are machine learning with geospatial, time series data, data-centric AI (AI4data), AI for simulation and decisions, multimodal AI and foundation models, LLM and agentic AI (routing, teaming, reasoning). He currently serves as an Associate Editor of ACM Transactions on Knowledge Discovery from Data.

Frontiers of Foundation Models for Time Series

Recent development in deep learning has spurred research advances in time series modeling and analysis. Practical applications of time series raise a series of new challenges in science applications, such as multi-resolution, multimodal, missing value, distributeness, and interpretability. In this talk, I will discuss possible paths to foundation models for time series data and future directions for time series research.

Biography:
Yan Liu is a full professor in the Computer Science Department of the Viterbi School of Engineering at the University of Southern California (USC), where she has been a faculty member since August 2010. Prior to joining USC, she worked as a research staff member in the Data Analytics Group at the IBM T.J. Watson Research Center from November 2006. She earned both her M.S. and Ph.D. degrees from Carnegie Mellon University. Her research focuses on machine learning and artificial intelligence, with applications spanning healthcare, sustainability—including climate science and traffic—and social media analysis. Her specific areas of interest include the development of novel machine learning models for time series analysis and modeling, explainable machine learning techniques, and efficient algorithms for sampling and optimization. Over the years, she has received numerous honors, including being named a 2021 ACM Distinguished Member and a 2020 National Academy of Inventors Senior Member. She was also recognized as a member of the 2018 New Voices of the National Academies of Science, Engineering, and Medicine and received the 2017 Biocom Catalyst Award. Additional accolades include the 2013 Okawa Foundation Research Award, the 2013 NSF CAREER Award, the 2011 James H. Zumberge Research and Innovation Fund Individual Award, the 2011 Yahoo Faculty Research & Engagement Award, an ACM Doctoral Dissertation Award Honorable Mention in 2007, and a Best Paper Award at the 2007 SIAM Conference on Data Mining.

Madhura Raut is a Principal Data Scientist at Workday, where she leads the design of large-scale machine learning systems for labor demand forecasting. Her work integrates time series modeling, AI, and reinforcement learning to enable real-time workforce optimization. She is the lead inventor on two U.S. patents related to advanced time series techniques, and her ML product has been recognized as a Top HR Product of the Year by Human Resource Executive. With over nine years of experience spanning enterprise machine learning and applied research, Madhura is passionate about building intelligent, production-grade systems that address complex operational challenges. Besides her professional work, she also serves as an IEEE Senior Member and mentor, actively contributing to the broader AI and data science community.

Veer Lade is a Senior Applied Scientist at UberEats leading the Item Intelligence team focusing on developing ranking and recommender solutions for the item ranking space. He has been the seed scientist in the team helping UberEats expand into the Grocery & New Verticals domain. His work focuses on Recommender model development, experimentation, developing ranking simulations and deriving key data insights to guide Business decisions. Veer is passionate about data foundations, objective functions and deploying production grade ML systems. .