基于集成森林元学习网络的客户流失预测

Customer churn prediction based on the integration of meta-learning network of the forest

为解决树模型在客户流失预测任务中较难捕捉时序特征的问题,提出了基于集成森林元学习网络(ensemble forest meta-learning network,EFML)的流失预测方法。首先通过分组等策略进行数据提质,并结合下采样技术解决样本类别不平衡问题。然后,基于EFML的语义图构造器构建用户时序特征的语义向量,以描绘用户细粒度行为,形成语义图并显式融合。最后,训练多个基础树模型作为元学习器-多层感知机(multilayer perceptron,MLP)的输入,生成综合的流失预测结果。实验证明,EFML能充分挖掘客户历史行为差异,捕获和学习基础树模型间的互补关系,相对于随机森林(random forest,RF),其AUC提升2.7%,AP提升3.7%,预测精度提升显著。该框架结合树模型和微观特征,具备卓越的解释性,为运营商实现更精细的用户化管理提供新视角。

To address the challenge of capturing temporal features in customer churn prediction tasks by tree models,a churn prediction method based on ensemble forest meta-learning network(EFML)was proposed.Firstly,data quality was improved through grouping strategies and class imbalance issues were addressed with undersampling techniques.Secondly,semantic vectors of user temporal features were constructed using EFML’s semantic graph constructor to depict fine-grained user behavior,forming a semantic graph and explicitly integrating it.Finally,multiple base tree models were trained as meta-learners,with the inputs being multilayer perceptron(MLP)to generate comprehensive churn prediction results.Experimental results demonstrate that EFML can effectively exploit differences in customer historical behaviors,capture and learn complementary relationships between base tree models.Compared to random forest(RF),EFML shows a 2.7%increase in AUC,a 3.7%increase in AP,and a significant improvement in prediction accuracy.This framework,combining tree models and micro-level features,possesses excellent interpretability,providing a new perspective for operators to achieve more refined user-centric management.