山东大学 林路教授：A race-DC in Big Data

2008年1月以来，Prof. Peter XK Song一直担任美国密西根大学公共健康学院生物统计系教授和研究副主任。1996年，他在加拿大温哥华的英属哥伦比亚大学获得统计学博士学位。自2015年以来，他在Nature, Science, AJHG, JASA, JRSSB, AOS, JMLR, JOE, Biometrika, Biometrics, JBES等国际顶级期刊上发表了66篇论文。他的研究兴趣包括分布推断，高维数据分析，纵向数据分析，缺失数据问题，时空建模，精确治疗的统计方法。他是 ASA的 Fellow和 International Statistical Institute的 Elected Member；现为JASA、 Canadian Journal of Statistics和 the Journal of Multivariate Analysis期刊的副主编。

内容提要：

Assessing intervention-led health benefit in practice is a notoriously hard problem due to many potential confounding factors, where multiple outcome variables are often used for the quantification of the benefit. Two major challenges arise from the setting of developing individual intervention allocation rules: (i) different outcomes are of different relevance to the underlying health benefit, and (ii) different sample sizes are available with different outcome variables. We propose a new support vector machine that allows us to synergize the heterogeneities across different outcomes through a weighted self-learning scheme. As an extension to the outcome weighted learning, the proposed synergistic self-learning incorporates both multiple outcomes and missing data in the optimization, leading to an optimal solution of individual rules for dietary intake with respect to a certain composite health benefit score. We establish the algorithmic convergence of the proposed synergistic self-learning, and illustrate the performance of the methodology through both simulation study and real data analysis of a calcium supplementation randomized trial. This is a joint work with Yiwang Zhou.

在实践中评估介入引导的健康益处是一个众所周知的困难问题，因为存在许多潜在的混杂因素，其中多个结果变量常常用于量化益处。制定个人干预分配规则的设置带来了两大挑战:(i)不同的结果与潜在的健康益处有不同的相关性，(ii)不同的样本量有不同的结果变量。我们提出了一种新的支持向量机，它允许我们通过加权的自学习方案来协同不同结果的异质性。作为结果加权学习的扩展，提出的协同自学习将多个结果和缺失的数据结合到优化中，在一定的综合健康效益评分下，得出饮食摄入的单个规则的最优解。我们建立了协同自学习的算法收敛性，并通过一个补钙随机试验的模拟研究和真实数据分析来说明该方法的性能。这是与Yiwang Zhou的合作文章。