Speaker: 孙法省，东北师范大学

Inviter: 何煦

Title: Model-free deterministic subsampling based on Kullback-Leibler divergence

Language: Chinese

Time & Venue: 2025.11.29 16:00-16:30 腾讯会议：979-347-585

Abstract: Subsampling is a useful approach in large-scale statistical learning. Most existing subsampling methods rely heavily on predefined models, making their performance highly sensitive to model misspecifications. To address this fundamental limitation, this paper introduces Kullback-Leibler subsampling, a novel model-free framework designed to ensure robustness across diverse modeling paradigms. The method uses the concept of generating Kullback-Leibler points, which is an efficient and robust technique for drawing samples from a specified continuous distribution. Theoretically, we show that the kernel density estimator of the selected subsample asymptotically converges to the true probability density function. A series of simulation studies and real-data experiments demonstrate that the proposed method is robust to diverse modeling approaches and outperforms several existing model-free subsampling methods in data compression tasks.