Holder： Xinwen Zhou（the University of Illinois Chicago）
Time：2026-12-10 14:00-15:00
Location：Siyuan Lecture Hall，Zhi Hua Building-225
Abstract:
Integrating information across related tasks can substantially improve estimation accuracy, but indiscriminate aggregation may deteriorate performance when some tasks are unrelated. This challenge is further complicated by heterogeneity among related tasks and differences in covariate distributions across tasks. In this work, we study nonparametric multitask regression in a reproducing kernel Hilbert space, which provides a flexible yet computationally tractable framework for modeling nonlinear relationships. We allow covariate distributions to vary across tasks and assume that an unknown majority of tasks have similar regression functions, while the remaining tasks may differ substantially.
We propose a robust multitask kernel ridge regression procedure that combines pooled estimation, task-specific residual correction, filtering-based robust aggregation, and adaptive personalization, with single-task estimators serving as safety anchors.
We establish finite-sample convergence rates that adapt to heterogeneity among the related tasks, while ensuring that every task retains, up to logarithmic factors, the convergence rate attainable by fitting that task separately. In particular, when all tasks share the same covariate distribution, we derive minimax lower bounds and show that our convergence rates match them up to logarithmic factors. Numerical experiments demonstrate substantial gains from information sharing among related tasks while maintaining reliable task-specific performance in the presence of unrelated tasks and covariate shift.
About the Speaker:
I am an Associate Professor in the Department of Information and Decision Sciences at the University of Illinois Chicago. Before joining UIC in 2023, I served as an Associate Professor in the Department of Mathematics at UC San Diego. My research centers on high-dimensional and robust statistics, quantile regression, and nonparametric methods. I also serve as an Associate Editor for the Journal of the Royal Statistical Society: Series B, Journal of the American Statistical Association, Statistica Sinica, and The Annals of Applied Probability.
Your participation is warmly welcomed!
欢迎扫码关注北大统计科学中心公众号，了解更多讲座信息!
