基于Group Lasso的多源电信数据离网用户分析

Group Lasso-Based Feature Selection for Off-network Analysis in Multisource Teledata

随着行业竞争愈演愈烈,电信企业的客户流失情况越来越严重,给电信企业造成了巨大损失.通过电信企业的数据来做离网用户的预测,从而进一步作出挽留客户的正确决策,成为电信企业日益关注的问题.面对电信后台汇总的多源数据,经分析发现其呈现天然的组结构.为了选择对于离网类别最具判别性的特征,本文使用了一种基于Group Lasso的组特征选择方法,在此基础上用交叉验证法选择适当的特征组,最终将选择出的少量组特征用于预测离网和停机的宽带用户.实验表明,在江苏某地级市电信离网用户分析数据中取得了比其他特征选择方法的精度平均高至少10%的预测性能.

With the intensified competition in the industry,customer churn analysis is becoming one of the most significant tasks for the telecom companies,which might lead great financial loss to them. Thus,using the data to predict potential off-network customers and then making business decisions to retain these customers,have drawn lots of attention nowadays. In this paper,we present a Group Lasso-based feature selection method to predict the latent off-network customers by analyzing the corresponding multisource teledata. Specifically,we utilize the cross-validation strategy to choose the optimal sets of feature groups. Extensive experiment results show that the proposed approach has the superior performance（ the Precision value is 10% higher than the other methods） on a real telecom dataset derived by a certain city in a prefectural city of Jiangsu.