基于蚁群算法的成本敏感线性集成多分类器的客户流失研究

针对电信客户流失预测问题的复杂性,融合自组织神经网络良好的连续属性值离散化优势、粗糙集理论出色的属性约简功能和蚁群优化算法全局的随机搜索特点,在模型集成技术和成本敏感学习理论的基础上,提出了一种新的基于蚁群算法的成本敏感线性集成多分类器的电信客户流失预测模型。构建该集成模型可分为4个阶段:(1)连续属性值的离散处理:利用自组织神经网络对连续属性值进行非监督离散化处理;(2)原始属性集的约简处理:使用粗糙集理论按属性重要性原则对离散属性进行约简;(3)子分类器的建立:分别使用NaiveBayes、Logistic回归、多层感知器和决策树等4种差异性很大的分类技术在约简属性集上建立4个对应的客户流失预测子分类器;(4)子分类器的集成:基于成本敏感学习理论,构建了4种不同的线性集成模型,采用蚁群优化算法求解集成模型的最优线性组合权重系数。将该模型应用于某电信客户流失预测,其实验结果表明该集成方法是可行且有效的。

Customer Churn Prediction Model Fusing Multiple Classifier Based on Cost Sensitivity Study Using Ant Colony Optimization

According to the complexity of customer churn prediction in Telecom,integrating the characters,such as the excellent indicative of the self-organizing neural network(SOM)when discretizing continuous at tributes,the outstanding capability of rough set theory(RS)when reducing the attributes,and the feature of the ant colony optimization(ACO)when searching atrandom globally,based on the technique of model integration and the theory of cost sensitivity study,a new customer churn model is proposed,i.e.fusing multiple classifiers based on cost sensitivity study using ant colony optimization(ACO). When constructing the model,there are four steps:(1)Discretizing the continuous at tributes unsupervis edly using SOM;(2)Reducing the discrete attributes according to the importance of the attributes using RS;(3)Building four sub classifiers on the reduced attributes sets using four completely different classification techniques including NaiveBayes,Logistic Regression,Multilayer Perceptron and Decision Tree,respectively;(4)Fusing the sub classifiers,that based on cost sensitive theory,and integrated four models linearly,which weight sobtained through the ant colony optimization(ACO).Applying the model to customer churn research in a telecom munication enterprise,the experiment results suggest that the fusing technique is feasible and very efficient.