Developing Optimization Models for Customer Churn Prediction in Telecoms using Artificial Intelligence Techniques

Customer churn prediction is a critical task in the telecommunication (telecoms) industry, where accurate identification of customers at risk of churning plays a vital role in reducing customer attrition. This research presents a comprehensive study on customer churn prediction using Machine Learning (ML) techniques. Three distinct aspects of churn prediction are investigated: class imbalance handling, feature selection, and model enhancement, each utilizing different AI methodologies. First, to address the issue of highly skewed datasets, we propose an effective oversampling method called HEOMGA. HEOMGA combines the Heterogeneous Euclidean-Overlap Metric (HEOM) and Genetic Algorithm (GA) to oversample the minority class. Experimental results on six benchmark datasets from the UCI repository demonstrate the superiority of HEOMGA over popular oversampling methods such as SMOTE, ADASYN, G SMOTE, and Gaussian oversampling, as evaluated by three performance metrics: Recall, G mean, and AUC. The experiment results show the effectiveness of the proposed method compared to some popular oversample methods, such as SMOTE, ADASYN, G SMOTE, and Gaussian oversampling methods. The HEOMGA method significantly outperformed the other oversampling methods in terms of recall, G mean, and AUC when the Wilcoxon signed-rank test is used. Second, in the preprocessing phase, feature selection plays a crucial role in improving the performance of ML models while reducing computational time. To address this, we introduce an ACO-RSA based-FS approach that combines two metaheuristic algorithms: Ant Colony Optimization (ACO) and Reptile Search Algorithm (RSA). The ACO-RSA approach selects the most salient features for churn prediction. The performance evaluations on six open-source customer churn prediction datasets demonstrate the superiority of ACO-RSA over other competitor algorithms such as PSO, MVO, GWO, standard ACO, and standard RSA. Third, we focus on enhancing Gradient Boosting Machine's (GBM) learning process for Churn Prediction (CP). In traditional GBM, learning process uses Decision Tree (DT) as a base learner and logistic loss as a loss function. However, using a DT to start the GBM model in the training process could result in poor predictive performance and overfitting. Therefore, a new model, called CP- Enhanced Gradient Boosting Model (CP- EGBM) is proposed. In the CP- EGBM, Support Vector Machine with a Radial Basis Function kernel (SVMRBF) is employed as a base learner and exponential loss function is utilized as a loss function to enhance learning process of the GBM. In order to effectively tune the hyperparameters of CP-EGBM, Finally, a modified version of Particle Swarm Optimization (PSO) using the consumption operator of the Artificial Ecosystem Optimization (AEO) method to prevent premature convergence of the PSO in the local optima is developed to tune the hyper-parameters of the CP-EGBM effectively. Six open-source CP datasets are used to evaluate the performance of the developed CP-EGBM model. Comparative analysis reveals the significant superiority of the CP-EGBM over GBM and SVM models, along with promising improvements compared to the recently reported models in the literature. Comparative analysis with state-of-the-art models showcases CP-EGBM's promising improvements, making it a robust and effective solution for churn prediction in the telecom industry. This research contributes to customer CP by providing effective solutions to address class imbalance, feature selection, and model enhancement challenges. The proposed methods, HEOMGA, ACO-RSA, and CP-EGBM, demonstrate their efficacy in improving CP performance, thereby assisting the telecom industry in understanding customer needs and taking appropriate actions to mitigate churn risks.