基于社交媒体数据的贝叶斯A/B检验  

Bayesian A/B Test Based on Social Media Data

作　　者：李薛莎 付英姿[1] 薛茜 夏思琴 LI Xue-sha;FU Ying-zi;XUE Xi;XIA Si-qin(Faculty of Science,Kunming University of Science and Technology,Kunming 650093,China)

机构地区：[1]昆明理工大学理学院,云南昆明650093

出　　处：《软件导刊》2021年第9期113-118,共6页Software Guide

基　　金：国家自然科学基金项目(11201200,11561035)。

摘　　要：A/B检验主要用于考察相对于原方案A,改进方案B是否更优。重点研究新旧版本网页点击率的贝叶斯A/B检验问题,通过建立起二项分布下的二元Logistic回归模型,结合拉普拉斯近似及重要性抽样技术,成功计算出边际似然并最终得到贝叶斯因子。而贝叶斯因子是贝叶斯A/B检验的核心,经典的A/B检验仅考虑A、B方案是否相等,基于此进一步考虑两者谁更优的问题。研究结果表明,对网页的改版并不能有效地增加用户点击率。The A/B test is mainly used to investigate whether the improved plan B is better than the original plan A.Focuses on the click-through rate by the Bayesian A/B test with the old and new versions of web pages,cauculate the marginal likelihood successfully by establishing a binary logistic regression model under the binomial distribution while combining Laplace approximation and importance sampling techniques,then the Bayes factor is finally obtained,The Bayes factor is the core of Bayesian A/B test.The classical A/B test only considers the equality of plans A and B,on this basis,this paper further discusses the question of whether A or B is better.The results of the research show that the revision of the web page does not effectively increase the click-through rate of users.

关 键 词：贝叶斯因子 A/B检验 模型选择 拉普拉斯近似 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]