报告题目：A Flexible Spatio-Temporal Stick-Breaking Count Regression

报告人：杜江

时间：10月17日 10：30-11：30

地点：国产自拍 106会议室

报告摘要：Covariate effects in spatio-temporal count data may exhibit complex patterns of local sharing across space and time. In this work, we propose a flexible Bayesian spatio-temporal stick-breaking count regression model that assigns spatio-temporally varying weights to globally shared coefficient atoms. This construction accommodates local sharing without imposing continuity on the coefficient process and admits an infinite-mixture representation that yields flexible predictive distributions. We derive explicit expressions for the prior co-clustering probability and induced response covariance, revealing how the kernel and knot distribution govern spatio-temporal dependence. The analysis further shows that common distance-decay kernels may limit attainable local dependence. A plateau kernel is therefore introduced, with a shape parameter adjusting the maximum local co-clustering strength and separate decay parameters controlling the spatial and temporal ranges. For posterior inference, we combine finite truncation with Pólya–Gamma augmentation to construct an efficient blocked Gibbs sampler, and show that the marginal approximation error induced by truncation decays geometrically. Simulations and an application to Spanish wildfire data demonstrate that STSBNB improves predictive performance relative to competing methods, while the proposed plateau kernel provides clear additional gains over conventional kernels, particularly under strongly localized dependence and discontinuous coefficient structures.

报告人简介：杜江，北京工业大学教授、博导。研究方向为函数型数据分析、空间数据分析、模型检验、高维数据分析、贝叶斯分析等。主持国家自然科学基金项目2项，中国博士后基金（面上）1项，北京市教委科技计划项目1项，参加国家重点研发计划2项、国家自然科学基金2项、国家社科科学基金2项。已在国内外学术刊物上发表论文60 余篇，其中50余篇被SCI检索。