基于分步特征提取和组合分类器的电信客户流失预测模型  

A telecom customer churn prediction model based on two-stage feature selection method and ensemble classifier

作　　者：徐子伟[1] 王传启[1] 王鹏[1] 黄海[1] 

机构地区：[1]中国科学技术大学信息科学技术学院,安徽合肥230027

出　　处：《微型机与应用》2016年第13期51-54,共4页Microcomputer & Its Applications

摘　　要：针对电信客户流失数据集存在的数据维度过高及单一分类器预测效果较弱的问题,结合过滤式和封装式特征选择方法的优点及组合分类器的较高预测能力,提出了一种基于Fisher比率与预测风险准则的分步特征选择方法结合组合分类器的电信客户流失预测模型。首先,基于Fisher比率从原始特征集合中提取具有较高判别能力的特征;其次,采用预测风险准则进一步选取对分类模型预测效果影响较大的特征;最后,构建基于平均概率输出和加权概率输出的组合分类器,以进一步提高客户流失预测效果。实验结果表明,相对于单步特征提取和单分类器模型,该方法能够提高对客户流失预测的效果。To solve the high dimensionality problem in telecom dataset and the weak forecasting ability of single classifiers,this paper proposes a telecom churn prediction model based on two-stage feature selection method and ensemble classifier,taking advantages of filter and wrapper selection method and ensemble classifiers with better forecasting performance.The two-stage feature selection method is based on Fisher＇s ratio and prediction risk.Firstly,features with high discriminative ability are selected by Fisher＇s ratio.Then we use prediction risk to further select features that have great impacts on classifiers.Lastly,two ensemble classifiers based on the average probability and weighted average probability are constructed to further improve the forecasting performance.Experimental results verify that the proposed method can improve the forecasting performance compared to the model based on one-step feature selection method or single classifier.

关 键 词：电信客户流失预测 分步特征提取 组合分类器 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]