COMPARATIVE ANALYSIS OF FEATURE SELECTION METHODS FOR CUSTOMER CHURN PREDICTION USING DEEP NEURAL NETWOR

Ezra Evanendra Harshaditya, Dr. Suprapto, M.I.Kom.

2025 | Skripsi | ILMU KOMPUTER

Customer churn poses a significant challenge for businesses, particularly in subscription-based industries such as telecommunications, where retaining customers is critical to maintaining profitability. Predicting and mitigating churn effectively requires advanced predictive modeling approaches. This research proposal focuses on investigating the impact of various feature selection techniques on customer churn prediction using Deep Neural Networks (DNNs). The study aims to compare the performance of four feature selection methods—Filter (ANOVA F-test), Wrapper (Random Forest), Dimensionality Reduction (PCA), and Embedded (XGBoost)—to understand their influence on model accuracy, computational efficiency, and interpretability. The research will utilize the Telco Customer Churn Dataset, which contains comprehensive customer data encompassing demographic, service-related, and account-specific attributes. The proposed study will systematically apply each feature selection method individually to train DNN models and evaluate their performance using metrics such as accuracy, precision, recall, F1-score, and Matthews Correlation Coefficient (MCC). A baseline model using all features will also be established for comparative analysis. The expected outcomes of this research include identifying the impact of feature selection method on DNN customer churn prediction model. The findings aim to provide actionable insights into the role of feature selection in optimizing churn prediction models, contributing to the development of scalable and accurate tools for customer retention strategies. This study seeks to advance the understanding of feature selection in high-dimensional datasets, offering valuable contributions to machine learning applications in customer analytics.

Customer churn poses a significant challenge for businesses, particularly in subscription-based industries such as telecommunications, where retaining customers is critical to maintaining profitability. Predicting and mitigating churn effectively requires advanced predictive modeling approaches. This research proposal focuses on investigating the impact of various feature selection techniques on customer churn prediction using Deep Neural Networks (DNNs). The study aims to compare the performance of four feature selection methods—Filter (ANOVA F-test), Wrapper (Random Forest), Dimensionality Reduction (PCA), and Embedded (XGBoost)—to understand their influence on model accuracy, computational efficiency, and interpretability. The research will utilize the Telco Customer Churn Dataset, which contains comprehensive customer data encompassing demographic, service-related, and account-specific attributes. The proposed study will systematically apply each feature selection method individually to train DNN models and evaluate their performance using metrics such as accuracy, precision, recall, F1-score, and Matthews Correlation Coefficient (MCC). A baseline model using all features will also be established for comparative analysis. The expected outcomes of this research include identifying the impact of feature selection method on DNN customer churn prediction model. The findings aim to provide actionable insights into the role of feature selection in optimizing churn prediction models, contributing to the development of scalable and accurate tools for customer retention strategies. This study seeks to advance the understanding of feature selection in high-dimensional datasets, offering valuable contributions to machine learning applications in customer analytics.

Kata Kunci : Customer Churn Prediction, Deep Neural Networks, Feature Selection, ANOVA F-test, Random Forest, PCA, XGBoost