报告题目：Generating tensor factor structure via diffusion model with Tucker Unet  

报 告 人：孔新兵教授 (东南大学)

报告时间：2025年12月30日（星期二）上午10:00

报告地点：广东工业大学龙洞校区行政楼610室

邀 请 人：韩晓卓

报告摘要: Large-dimensional factor models are widely used in asset pricing, macro economic analysis, international trading, social network analysis, machine learning, and so on. In this talk, we first give a new statistical testing for the existence of the tensor factor structure in large dimension. We show that our test is powerful which is theoretically underpinned compared with the most recent works. Empirically, we found that the structure is significant for the MNIST hand written numbers while not that significant for the two-way arranged return matrix series. Motivated by the data augmentation, we studied how to generate tensors with low-rank factors via the diffusion generative model with our carefully tailored Tucker Unet. We theoretically present the accuracy of the deep generative artifically networks, including the approximation error, estimation error and the TV error of the large-dimensional distribution function of the generated tensors. Empirical studies in picture generation and Newyork taxi data show the good performance of our Tucker Unet.

报告人简介：孔新兵，现为东南大学统计与数据科学学院教授。主要研究方向为高维统计，计量经济学和机器学习。在统计学和计量经济学国际顶级期刊AoS, JASA, Biometrika, JoE, JBES发表论文20余篇，其中在AoS, Biometrika独立发表论文3篇，JoE 近三年most popular一篇，JBES近三年 Most cited一篇。主持国家自然科学基金项目5项，其中重点专项1项。联合主持中钢集团CAD对象识别和自动算量系统项目和招商银行南京分行大模型应用项目。获第一届统计科学技术进步奖一等奖，江苏省教育系统先进个人（优秀教师），完成省教育教学改革重点项目1项，参编统计学101教材《统计机器学习》。担任JBES期刊编委。