Intelligent churn prediction in telecom: employing mRMR feature selection and RotBoost based ensemble classification
Intelligent churn prediction in telecom: employing mRMR feature selection and RotBoost based ensemble classification
复制标题
DOI：
10.1007/s10489-013-0440-x
复制
发表时间：
2013-10
影响因子：
5.3
通讯作者：
Adnan Idris;Asifullah Khan;Yeon Soo Lee
中科院分区：
文献类型：
--
作者：
Adnan Idris;Asifullah Khan;Yeon Soo Lee
由于电信数据集的海量性质，电信流失预测最近引起了利益相关者的极大兴趣，预测电信流失是一个具有挑战性的问题。在这方面，我们利用高效的特征提取技术和集成方法，提出了一种面向电信行业的智能流失预测系统。我们使用了随机森林、旋转森林、RotBoost和装饰集成，并结合最小冗余度和最大相关性(MRMR)、Fisher‘s Ratio和F-Score方法对电信流失预测问题进行了建模。我们观察到，与Fisher比率和F-Score相比，mRMR方法返回了最具解释性的特征，这显著减少了计算量，并有助于集成获得更好的性能。与随机森林、轮换森林和装饰相比，在标准电信数据集上，结合mRMR特征的RotBoost获得了更好的预测性能。RotBoost集成的较好性能在很大程度上归功于特征空间的旋转，这使得基分类器能够学习搅拌者和非搅拌者的不同方面。此外，RotBoost中的Adbooing过程也有助于通过处理硬实例来实现更高的预测精度。在标准电信数据集上使用AUC、基于灵敏度和基于特异度的度量进行性能评估。仿真结果表明，基于RotBoost和mRMR特征相结合的方法(CP-MRB)能够有效地处理电信数据集的高维数据。CP-MRB在预测客户流失方面具有较高的准确率，因此在电信客户流失预测这一具有挑战性的问题上具有很好的建模前景。
Churn prediction in telecom has recently gained substantial interest of stakeholders because of associated revenue losses.Predicting telecom churners, is a challenging problem due to the enormous nature of the telecom datasets. In this regard, we propose an intelligent churn prediction system for telecom by employing efficient feature extraction technique and ensemble method. We have used Random Forest, Rotation Forest, RotBoost and DECORATE ensembles in combination with minimum redundancy and maximum relevance (mRMR), Fisher’s ratio and F-score methods to model the telecom churn prediction problem. We have observed that mRMR method returns most explanatory features compared to Fisher’s ratio and F-score, which significantly reduces the computations and help ensembles in attaining improved performance. In comparison to Random Forest, Rotation Forest and DECORATE, RotBoost in combination with mRMR features attains better prediction performance on the standard telecom datasets. The better performance of RotBoost ensemble is largely attributed to the rotation of feature space, which enables the base classifier to learn different aspects of the churners and non-churners. Moreover, the Adaboosting process in RotBoost also contributes in achieving higher prediction accuracy by handling hard instances. The performance evaluation is conducted on standard telecom datasets using AUC, sensitivity and specificity based measures. Simulation results reveal that the proposed approach based on RotBoost in combination with mRMR features (CP-MRB) is effective in handling high dimensionality of the telecom datasets. CP-MRB offers higher accuracy in predicting churners and thus is quite prospective in modeling the challenging problems of customer churn prediction in telecom.