主 题：Prediction Models for Network-linked Data
主讲人：朱冀美国密歇根大学统计系教授
主持人：石磊统计与数学学院院长
时 间：2016年12月26日（周一）上午10:00-11:00
地 点：北院卓远楼305
主办单位：统计与数学学院
摘 要：Prediction problems typically assume the training data areindependent samples, but in many modern applications samples come fromindividuals connected by a network. For example, in adolescent healthstudies of risk-taking behaviors, information on the subjects' socialnetworks is often available and plays an important role throughnetwork cohesion, the empirically observed phenomenon of friendsbehaving similarly. Taking cohesion into account in prediction modelsshould allow us to improve their performance. Here we propose aregression model with a network-based penalty on individual nodeeffects to encourage similarity between predictions for linked nodes,and show that it performs better than traditional models boththeoretically and empirically when network cohesion is present. Theframework is easily extended to other models, such as the generalizedlinear model and Cox's proportional hazard model. Applications topredicting levels of recreational activity and marijuana usage amongteenagers based on both demographic covariates and their friendship
networks are discussed in detail and demonstrate the effectiveness ofour approach.
朱冀教授简介：美国斯坦福大学统计学博士，美国密歇根大学统计系教授，研究领域为统计机器学习与数据挖掘、研究兴趣包括高维数据分析、网络数据分析等，在国际主流学术刊物上共发表70 多篇学术论文，担任包括国际统计学顶尖刊物《Journal of the American Statistical Association》、《Biometrika》在内的多个期刊副主编。
