时 间：2023年12月1日14:30-15:30
地 点：普陀校区理科大楼A1514
报告人：郁文 复旦大学教授
主持人：马慧娟 华东师范大学副教授
摘 要：
We propose a powerful and flexible neural modeling framework for survival regression. The framework basically assumes a separated structure of the baseline hazard rate and the nonlinear covariates effect. Meanwhile, a multiplicative frailty is introduced to capture the unobserved heterogeneity among individuals and the deep neural network architectures are adopted to approximate the baseline hazard rate and the nonlinear covariate structures, leading the proposed framework called neural frailty machines (NFM). The NFM can be viewed as an extension of neural proportional hazard models and includes many commonly used survival regression models as special cases. The likelihood function for right censored data is used to serve as the objective. The proposed algorithm allows efficient stochastic training, which can easily scale to large datasets. The estimation accuracy is measured by a metric defined through a Hellinger-type distance for hazard rate function. The non-asymptotic bounds for the estimation errors based on the Hellinger-type distance are derived. Then the consistency of the proposed neural estimators is established and the convergence rates are obtained. The rates are shown to reach the optimal speed of nonparametric regression estimation. Some simulation studies are carried out to verify the theoretical findings. The prediction performance of the proposed NFM models is evaluated over 6 benchmark datasets with different scales. The results show evidence on the improvement of the proposed method compared with the existing state-of-the-art survival models.
报告人简介：
郁文，复旦大学管理学院统计与数据科学系教授、博士生导师，主要从事生存分析、半参数模型、两阶段抽样设计、经验似然等方向的研究，在JRSSB、JASA、中国科学等国内外学术期刊以及NeurIPS等国际会议上发表学术论文三十余篇，主持国家自然科学基金面上项目、青年项目以及教育部博士点基金等研究工作。