Optimizing Artificial Neural Network for Customer Churn: Advanced Data Balancing and Feature Selection

Customers are valuable assets in the dynamic business world. However, service dissatisfaction often leads them to switch to competitors, a phenomenon known as customer churn. In the telecommunications industry, churn poses a significant challenge as it directly impacts revenue and influences other customers within their social networks to do the same. Consequently, predicting churn has become essential, with numerous researchers employing various methods to classify potential churners. This study builds upon prior research that utilized Artificial Neural Networks (ANN) or Deep Learning to predict churn, achieving an accuracy of 88.12%. To improve model performance, this research implements an Artificial Neural Network (ANN) as the primary algorithm, along with Random Over-Sampling (ROS) and Synthetic Minority Oversampling Technique (SMOTE) for data balancing, and three feature selection methods: Minimum Redundancy Maximum Relevance (mRMR), Lasso Regression, and XGBoost. The results demonstrate a 0.38% increase in accuracy compared to previous studies. The finding suggests opportunities for further exploration. Future studies can consider alternative feature selection techniques, such as Wrapper Methods or Heuristic/Metaheuristic approaches, which may produce more optimal feature combinations. Other data balancing methods, such as Undersampling techniques (e.g., Random Undersampling, Tomek Links) or Hybrid Methods (e.g., SMOTE combined with Tomek Links), could be explored to address imbalanced datasets effectively. These approaches are expected to provide better combinations and to improve overall prediction performance, enabling researchers to develop more robust and accurate models for customer churn prediction in subsequent studies.