学术报告20250510：An efficient Bayesian design method for estimating Shapley values

报告题目：An efficient Bayesian design method for estimating Shapley values

报告人：周正 讲师 北京工业大学

报告时间：2025年5月10日9:00-10:00

报告地点：腾讯会议838349331

摘 要：The Shapley value is a well-known concept in cooperative game theory, which provides a fair way to allocate cooperative gains (or costs) to players. However, computing Shapley values for a game involving d players requires evaluating values of all 2^d player coalitions, which is infeasible for a large d. Recently, the Shapley value has found widespread application in artificial intelligence, data science and many other emerging fields. In problems within these domains, calculating the value of a single coalition can be quite costly, further increasing the difficulty of obtaining Shapley values. In response to the computational issue, this paper proposes an efficient design-based method for estimating Shapley values. This method provides a prior distribution for values of a cooperative game, and then selects the design formed by key coalitions to provide Bayesian estimations of Shapley values by minimizing their posterior variances. The proposed method has been theoretically proven to ensure fairness for every player. Moreover, theoretical results are developed to substantially decrease the complexity in design and estimation processes. Compared with existing methods, this method offers three key advantages. First, it only requires computing the values of at least d^2-d+1 different coalitions to provide precise estimates of Shapley values for d players. Second, it allows for statistical inference of the estimates, so the confidence intervals of Shapley values can be provided. Third, it imposes no restrictions on the types of cooperative games, exhibiting strong robustness. Multiple case studies demonstrate that, our method outperforms existing methods in terms of estimation accuracy and identifying key players, at the same or lower costs.

报告人简介：周正，北京工业大学统计与数据科学系讲师。他于2024年在南开大学取得统计学博士学位，曾于2022至2023年前往美国田纳西大学进行学术访问。他的研究方向是试验设计，也包括其在数据科学和机器学习中的应用。他在Biometrika, Technometrics, IEEE TKDE等高水平学术期刊上发表了共计5篇文章。

欢迎广大师生参加！

内蒙古大学数学科学学院

2025年5月7日