模糊树节点的随机森林与异常点检测  

A random forest algorithm based on fuzzy tree node for anomaly detection

作　　者：胡淼 王开军 李海超[1,2] 陈黎飞 Hu Miao;Wang Kaijun;Li Haichao;Chen Lifei(College of Mathematics and Informatics,Fuiian Normal University,Fuzhou,350117,China;Digit Fujian Internet-of-Things Laboratory of Environmental Monitoring,Fujian Normal University, Fuzhou,350117,China)

机构地区：[1]福建师范大学数学与信息学院,福州350117 [2]数字福建环境监测物联网实验室,福建师范大学,福州350117

出　　处：《南京大学学报（自然科学版）》2018年第6期1141-1151,共11页Journal of Nanjing University（Natural Science）

基　　金：国家自然科学基金(61672157);福建省自然科学基金(2018J01778)

摘　　要：提出一种模糊树节点的随机森林算法进行异常点检测.在构建随机森林的分类决策树过程中,把模糊方法引入到二叉决策树的节点中,在节点中设计关于类别划分的模糊区域,在模糊区域上设计正常与异常隶属度函数.当某样本通过决策树节点的模糊区域时,若该样本的异常隶属度大于正常隶属度,则该样本被判别为异常类;否则,该样本进入决策树的下层树节点,若无下层节点则被判别为正常类.该样本的最终类别由随机森林算法中的投票步骤决定.在四个UCI数据集上的实验结果表明,在异常点检测的综合性能(召回率、精度和准确率)上,与基于随机森林的异常点检测算法RFV和RFP相比,新方法不仅具有较高的综合性能且性能稳定,还具有与一类支持向量机相当的性能,其部分实验结果优于一类支持向量机.This paper proposes a random forest algorithm based on fuzzy tree node for anomaly detection.In the process of constructing the classification tree of the random forests,the fuzzy method is introduced into the nodes of the binary decision tree.The fuzzy regions about the class division are designed in the nodes,and the normal and anomaly membership functions are designed on the fuzzy regions.When a sample passes through the fuzzy region of the decision tree node,if the sample’s anomaly membership degree is greater than the normal membership degree,the sample is discriminated as the anomaly class.Otherwise,the sample enters the lower tree node of the decision tree and can be identified as a normal class if there is no lower node.The final class of the sample is determined by the voting steps in the random forest algorithm.The experimental results on four UCI datasets show that on the overallperformance of the anomaly detection(recall,precision and accuracy).Compared with the RFV and RFP based on the random forest anomaly detection algorithm,the new method not only has higher comprehensive capability but also is stable,while RFV and RFP have lower performance in most cases.Compared with One-class Support Vector Machines,the new method has the same performance,and some of its experimental results are superior to One-class Support Vector Machines.

关 键 词：异常点检测 集成学习 随机森林 模糊隶属度函数 模糊树节点 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]