A Learning Method for Class Imbalance Problem: A Case Study of Churn Prediction

Abstract

The issue of class imbalance, particularly in relation to Machine Learning (ML) models, are common challenge in practical applications. ML is a field in which class imbalance is problematic, as it slows down the best learning process even by the best ML models. This issue significantly impacts performance, as these methods often prioritize learning the majority class while neglecting the distribution of the minority classes. In this paper, a novel oversampling method based on the stratification of Pascal's triangle is introduced to tackle the problem of class imbalance. The method is designed to enhance the learning process of ML models by facilitating a more effective representation of minority classes.  To evaluate effectiveness of the developed method, we conduct experiments on six benchmark datasets for the application of churn prediction. The results indicate that the developed method consistently outperforms SMOTE, ADASYN, G-SMOTE, and Gaussian oversampling techniques. Also, this approach appears to be not only an effective but also an efficient option for improving ML models' learning processes under the conditions of class imbalance.