题目：A Fusion penalized logistic threshold regression model with application to diabetes prognosis
报告人：
尹建鑫，中国人民大学统计学院副教授，副院长，博士生导师。2009年在北京大学数学科学学院获得博士学位。2009年至2011年在美国宾夕法尼亚大学医学院生物统计系做博士后研究。 2011年8月回国到中国人民大学任教。从事高维数据分析、图模型学习、文本分析、自适应实验设计、空间统计、非参数统计等方面的研究。研究成果大多发表在国际知名统计、科学杂志上（AOAS，JMVA，Stat Sinica，Sci Rep）。社会兼职包括中国现场统计研究会监事、计算统计分会副理事长、数据科学与人工智能分会（筹）副理事长、高维数据统计分会理事，中国人工智能学会不确定性人工智能专委会委员，国际生物统计学会中国分会副秘书长、北京生物医学统计与数据管理研究会理事等。全国统计教材编审委员第七届委员会专业委员。并曾作为参赛队代表获因果与预测国际挑战赛“最佳整体贡献奖”（2008年）。2015年获教育部第七届高等学校科学研究优秀成果奖（人文社会科学）统计学三等奖。
报告摘要：
In last decades, diabetes has become one of the major public health issues and its early prognosis has drawn considerable attention. Among a variety of practices, utilizing the physical examination data to develop a risk score for diabetes prognosis is popular given its affordability. However, it is common that different cohorts may have differ-ent patterns in physiology, which suggests a nonlinear relationship between the risk fa-ctors and the scores. Specifically, the prognosis may completely alter when certain variable(s) reaches some unknown threshold points, potentially of multiple levels. On the other hand, the risk score is expected to vary continuously with respect to some explanatory variable(s). That is, the risk score should behave similarly across adjacent levels of that variable(s). Traditional models, such as the logistic regression or tree- based methods, fail to account for these critical subtleties. To address these challenges, we propose a FusIon penalized Logistic ThrEshold Regression model (FILTER), where the fused Lasso reinforces the continuity of risk score with respect to the conti-nuous risk factors. We develop a tree-based estimator of the threshold points and obta-in the non-asymptotic rate of convergence. For the regression coefficients, estimation consistency and the sign consistency for selections are both derived. These statistical guarantees are further validated by comprehensive numerical studies, which demonst-rate the good performance of the proposedmethod. We also apply the FILTER model to a physical examination data to demonstrate building risk scores for diabetes prog-nosis.
报告时间：
2019年4月22日，周一，14:30-16:00
报告地点：
统计学院办公楼105会议室
