马尔可夫毯与多数投票因果发现评估  

Markov Blanket and Majority Voting for Causal Discovery Assessment

作　　者：李廷鹏 王雷[2] 彭丹华 廖军[2] 刘礼 LI Ting-peng;WANG Lei;PENG Dan-hua;LIAO Jun;LIU Li(State Key Laboratory of Complex Electromagnetic Environmental Effects on Electronics and Information System,Luoyang 471003,China;School of Big Data and Software Engineering,Chongqing University,Chongqing 401331,China)

机构地区：[1]电子信息系统复杂电磁环境效应国家重点实验室,河南洛阳471003 [2]重庆大学大数据与软件学院,重庆401331

出　　处：《计算机技术与发展》2025年第3期117-124,共8页Computer Technology and Development

基　　金：电子信息系统复杂电磁环境效应国家重点实验室项目(CEMEE2023G0202);国家自然科学基金(62207007);国家重大研发计划(2022YFB3303302)。

摘　　要：近年来,因果学习因其卓越的可解释性,成功地与深度学习相结合。在因果学习中,由于自然数据的收集难度和高成本,过去的研究主要依赖于合成数据集进行因果发现的验证。然而,合成数据集和半真实数据集常包含较多的人工控制,无法真实反映因果发现算法在实际场景中的表现。为解决这一问题,提出了一种在缺乏真实因果图的情况下评估因果发现方法的新策略。具体而言,将数据集划分为训练集和测试集,在训练集上进行因果发现以构建因果图,然后在测试集上验证该因果图。验证过程包括马尔可夫毯测试和因果图中每条边的因果方向判别,最终通过多数投票策略集成判别结果。在合成数据集和真实数据集上进行了全面的实验,结果表明,该方法在评估因果图的准确性和泛化性方面具有显著的有效性。这一方法为因果发现算法在真实场景中的性能评估提供了新的途径,提升了因果学习的应用潜力和可信度。In recent years,causal learning has successfully merged with deep learning due to its excellent interpretability.In causal learning,the collection of natural data is often difficult and costly,leading past research to primarily rely on synthetic datasets for validating causal discovery.However,synthetic and semi-real datasets often involve significant artificial control and fail to accurately reflect the performance of causal discovery algorithms in real-world scenarios.To address this issue,we propose a novel strategy for evaluating causal discovery methods in the absence of true causal graphs.Specifically,we divide the dataset into training and testing sets,perform causal discovery on the training set to construct a causal graph,and then validate this causal graph using the testing set.The validation process includes Markov blanket tests and causal direction identification for each edge in the causal graph,with the final results integrated using a majority voting strategy.We conducted extensive experiments on both synthetic and real datasets,and the results demonstrate that the proposed method effectively evaluates the accuracy and generalizability of causal graphs.The proposed method provides a new approach for assessing the performance of causal discovery algorithms in real-world settings,enhancing the applicability and reliability of causal learning.

关 键 词：因果发现 马尔可夫毯测试 数据集分割 多数投票策略 因果非对称识别方法 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]