基于专家知识的主动因果效应辨识  

Active causal effect identification with expert knowledge

作　　者：王天佐 周志华[1,2] Tian-Zuo WANG;Zhi-Hua ZHOU(National Key Laboratory for Novel Software Technology,Nanjing University,Nanjing 210023,China;School of Artificial Intelligence,Nanjing University,Nanjing 210023,China)

机构地区：[1]南京大学计算机软件新技术全国重点实验室,南京210023 [2]南京大学人工智能学院,南京210023

出　　处：《中国科学：信息科学》2023年第12期2341-2354,共14页Scientia Sinica(Informationis)

基　　金：科技创新2030“新一代人工智能”重大项目(批准号:2022ZD0114800);软件新技术与产业化协同创新中心资助项目。

摘　　要：因果效应定量描述变量间的影响作用,在许多现实任务中发挥着重要作用.辨识因果效应需要依赖已知的因果关系.一种确定因果关系的主流方法是引入干预信息,即通过人为干预某些因素而获取的信息.鉴于实际干预通常会带来巨大的成本,本文提出利用易于获取的专家知识替代干预,通过设计形如“在XX干预下,目标变量的值相较不干预时增大还是减小?”的问题并向专家咨询,确定因果关系,进而辨识因果效应.尽管每次专家返回的信息量少于干预信息,但通过一定次数的咨询,本文方法可以达到与使用干预信息方法相近的因果效应辨识效果.实验验证了上述结论.Causal effects play a crucial role in quantifying the impact between variables and are widely applied to various real-world problems.The identification of causal effects relies on known causal relations.To learn causal relations,traditional methods incorporate interventional information obtained by manipulating factors.However,such interventions can be costly in practice.Our study introduces an alternative method leveraging expert knowledge instead of interventions.Our methodology involves formulating questions such as“does the value of the target variable increase or decrease under intervention?”and posing these queries to experts.This allows our approach to learn causal relations,facilitating causal effect identification.Although less information than interventional information is provided by each expert consultation,our method achieves comparable results in causal effect identification through a sufficient number of inquiries.Experimental results validate this finding.

关 键 词：因果效应辨识 因果关系 人机交互 专家知识 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]