Identification of Customer Churn Considering Difficult Case Mining

In the process of user churn modeling, due to imbalance between lost users and retained users, use traditional classification models often cannot accurately comprehensively identify with tendency. To address this issue, it is not sufficient simply increase misclassification cost minority class samples in cost-sensitive methods. This paper proposes using Focal Loss hard example mining technique add weight α focus parameter γ cross-entropy loss function LightGBM. addition, emphasizes identification customers at risk churning raises for difficult-to-classify samples. On basis preceding ideas, FocalLoss_LightGBM model proposed, along random forests, SVM, XGBoost, Empirical analysis based on a dataset credit card publicly available Kaggle website. The AUC, TPR, G-mean index values were superior existing model, which can effectively improve accuracy stability potential users.