基于不平衡数据集的客户流失预测研究  

Customer churn prediction in class imbalance dataset

作　　者：杨智[1] 夏国恩[2] 金炜东[1] 

机构地区：[1]西南交通大学电气工程学院,成都610031 [2]广西财经学院,南宁530003

出　　处：《计算机应用研究》2010年第12期4447-4449,4453,共4页Application Research of Computers

基　　金：国家自然科学基金资助项目(70801021);国家教育部人文社会科学研究基金资助项目(08JC630019);广西高校优秀人才计划资助项目(桂教人[2009]62号文)

摘　　要：目前客户流失预测面临的主要问题之一就是类不平衡性(class imbalance)。针对这个问题,首先应用欠抽样法(undersampling)处理客户流失数据降低不平衡性,再应用C4.5D、C4.5N、RIPPER、NaiveBayes和Ran-domForest机器学习方法对客户流失进行预测。实验结果表明,欠抽样法是在牺牲负类样本预测精度的前提下,提高正类预测精度,于是采用重复抽样法(resampling)来弥补欠抽样法的缺陷,减少负类样本中含有大量有用信息的丢失,实验结果证明了这种方法的正确性和有效性。Class imbalance is one of main obstacles in customer churn prediction. To overcome this problem,this paper applied undersampling technology to reducing class imbalance,and applied 5 classifiers,such as C4. 5D,C4. 5N,RIPPER,NaiveBayes and RandomForest,to predicting customer churn. Undersampling technology increased predictive accuracy of customer churn by sacrificing predictive accuracy of normal customer,according to experimental results. So,applied resample technology to preprocessing customer churn dataset. Proved resample technology to make up shortage of undersampling techno-logy. So,resample technology is more suitable for customer churn dataset,which have characteristic of class imbalance.

关 键 词：类不平衡性 客户流失预测 机器学习 抽样法 

分 类 号：TP182[自动化与计算机技术—控制理论与控制工程]