Abstract
Customer churn prediction remains a critical concern for contemporary enterprises, given that the retention of existing clientele is substantially more cost-efficient than the acquisition of new customers. Accurate identification of potential churners facilitates the formulation of targeted retention strategies and enhances customer relationship management practices. This research investigates the application of machine learning algorithms to churn prediction, with a particular emphasis on addressing the prevalent issue of class imbalance in churn-related datasets. A novel Ratio-Based Data Balancing technique is introduced as a preprocessing mechanism to alleviate class distribution skewness and augment model efficacy. The proposed approach is systematically evaluated against conventional resampling methodologies using a diverse set of classifiers, including Support Vector Machine (SVM), Gradient Boosting, and Random Forest. Empirical findings indicate that the Random Forest classifier, when applied to datasets balanced via the proposed Ratio-Based method, consistently yields superior performance in terms of accuracy, precision, recall, and F1-score across various experimental conditions. Importantly, the method demonstrates marked improvements over traditional oversampling and undersampling techniques. The study substantiates that the integration of robust ensemble learning algorithms with an effective data balancing strategy substantially enhances predictive reliability and scalability. These results underscore the efficacy of the Random Forest algorithm, in conjunction with Ratio-Based balancing, as a viable framework for accurate and scalable customer churn prediction.

Keywords: Customer churn, Machine learning, Data Balancing, Prediction, Machine Learning Algorithm, Data resampling


Objectives
The principal aim of this study is to devise an accurate and efficient machine learning-based framework for customer churn prediction by addressing the inherent challenges associated with class imbalance in churn datasets. The specific objectives of this research are delineated as follows:
1. To introduce a novel Ratio-Based Data Balancing technique that alleviates class skewness and augments the efficacy of classification models by preserving an optimal class distribution during the data preprocessing phase.
2. To assess the performance of various machine learning algorithms—namely, Random Forest, Gradient Boosting, and Support Vector Machine (SVM)—on both imbalanced and balanced datasets, with a particular emphasis on their proficiency in accurately forecasting customer churn.
3. To undertake a comparative evaluation of conventional resampling strategies, such as oversampling and undersampling, vis-à-vis the proposed Ratio-Based approach in enhancing predictive metrics including accuracy, precision, recall, and F1-score.
4. To determine the most effective model among the selected classifiers, establishing that the Random Forest algorithm delivers superior predictive capability when trained on Ratio-Balanced data, thereby rendering it exceptionally well-suited to churn prediction applications.
5. To elucidate the practical ramifications of improved churn prediction accuracy for enterprises, encompassing more judicious resource allocation, tailored customer retention initiatives, and a consequent reduction in customer acquisition expenditures.