吃瓜51

报告时间：2025年12月19日10:00

报告地点：吃瓜51 4107

报告名称：Blue Noise-based Generative Models for the Imputation of Time-Series Data

报告摘要：Missing data imputation remains a critical challenge in highdimensional time-series data analysis, where traditional methods oftenstruggle to capture complex nonlinear dependencies inherent in sequentialdata. Diffusion-based generative models have shown state-of-the-artperformance by modeling the conditional distribution of missing valuesgiven observed data. However, these models typically rely on isotropicwhite noise during training, which can obscure important frequencydependent correlations that are crucial for accurate imputation. Toaddress the limitations of conventional imputation methods, we propose anovel approach called time-varying blue noise-based conditional scorebased diffusion model (tBN-CSDI). By modulating the noise scheduleaccording to the frequency characteristics of the data, tBN-CSDI improvesthe recovery of subtle, high-frequency temporal patterns that are oftenoverlooked by existing techniques. Experimental results on highdimensional datasets show that tBN-CSDI consistently outperforms existingimputation methods, achieving over a 30% reduction in imputation errorunder high data sparsity.

专家简介：宫海军博士于2009年在美国卡内基梅隆大学（Carnegie Mellon University,CMU）获得博士学位，师从著名计算生物学家Russell Schwartz教授，博士论文聚焦于蛋白质转运过程的建模与仿真。随后，他在卡内基梅隆大学计算机科学系从事博士后研究，合作导师为图灵奖得主Edmund M. Clarke教授，致力于将模型检测（Model Checking）技术应用于细胞信号通路的分析。自2012年起，宫海军博士加入圣路易斯大学（Saint Louis University）数学与统计系，现任统计学教授。他的研究方向主要集中在计算系统生物学、机器学习与形式化方法在生命科学中的应用，基因调控网络建模,单细胞数据分析, 缺失值填补方法，相关工作受到美国国立卫生研究院（NIH）的持续资助。