基于因果效应的K2算法的研究与改进  

Research and Improvement of K2 Algorithm Based on Casual Effect

作　　者：谈子健 屈红冰 肖寒 张捷[1] 

机构地区：[1]南京理工大学自动化学院,江苏南京210094

出　　处：《工业控制计算机》2024年第7期116-117,120,共3页Industrial Control Computer

摘　　要：贝叶斯网络是在不确定环境中进行推理的重要模型。从数据中正确学习贝叶斯网络结构是研究贝叶斯网络的重难点。为解决K2算法的学习效果受到输入节点顺序影响的问题,基于Pearl的因果理论,提出了一种基于因果效应的贝叶斯网络结构学习算法。该算法首先利用定义的因果效应强度来学习网络节点顺序;其次将获得的网络节点顺序输入K2算法得到初始网络结构;最后通过定义的因果效应强度删除边来修正初始网络结构。在贝叶斯网络标准数据集Asia和Alarm上的实验表明该方法对小型和大型网络都具有较好的学习效果。Bayesian network is an important model for reasoning in uncertain environments.Learning the structure of Bayesian networks correctly from data is a challenging task.To address the issue of the learning effectiveness of the K2 algorithm being influenced by the order of input nodes,a causal-effect-based Bayesian network structure learning algorithm is proposed based on Pearl's causal theory.Firstly,this algorithm learns the network node order using defined causal effect strengths.Secondly,the obtained network node order is input into the K2 algorithm to obtain an initial network structure.Finally,the initial network structure is modified by removing edges based on the defined causal effect strengths.The experiments on the standard datasets Asia and Alarm for Bayesian networks demonstrate that this method exhibits good learning performance for both small-scale and large-scale networks.

关 键 词：贝叶斯网络 K2算法 因果效应 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]