A Hybrid System for Customer Churn Prediction and Retention Analysis via Supervised Learning

Abstract

Telecom industry relies on churn prediction models to retain their customers. These prediction models help in precise and right time recognition of future switching by a group of customers to other service providers. Retention not only contributes to the profit of an organization, but it is also important for upholding a position in the competitive market. In the past, numerous churn prediction models have been proposed, but the current models have a number of flaws that prevent them from being used in real-world largescale telecom datasets. These schemes, fail to incorporate frequently changing requirements. Data sparsity, noisy data, and the imbalanced nature of the dataset are the other main challenges for an accurate prediction. In this paper, we propose a hybrid model, name as "A Hybrid System for Customer Churn Prediction and Retention Analysis via Supervised Learning (HCPRs)" that used Synthetic Minority Over-Sampling Technique (SMOTE) and Particle Swarm Optimization (PSO) to address the issue of imbalance class data and feature selection. Data cleaning and normalization has been done on big Orange dataset contains 15000 features along with 50000 entities. Substantial experiments are performed to test and validate the model on Random Forest (RF), Linear Regression (LR), Naïve Bayes (NB) and XG-Boost. Results show that the proposed model when used with XGBoost classifier, has greater Accuracy Under Curve (AUC) of 98% as compared with other methods.

The causes of bank customer churn based on XGBoost and LightGBMmodels: the evidence from the Kaggle dataset

“…Thakkar et al [30] proposed a novel class-dependent costsensitive boosting algorithm called AdaboostWithCost that reduces errors and misclassification costs, thereby reducing the cost of churn. Arshad et al [31] on the other hand, proposed a hybrid model called "A Hybrid System for Customer Churn Prediction and Retention Analysis via Supervised Learning (HCPRs)" which used Synthetic Minority Over-Sampling Technique (SMOTE) and Particle Swarm Optimization (PSO) to solve the problem of imbalance class data and feature selection. Extensive experiments are conducted to assess the model's performance across Random Forest (RF), Linear Regression (LR), Naïve Bayes (NB), and XGBoost.…”

“…Both XGBoost and LightGBM are gradient-boostingbased decision tree integration algorithms commonly employed to address classification and regression issues [31]. However, they differ in their splitting strategy.…”

“…Despite their different approaches, they both employ a gradient-boosting framework, which can promote weak and strong learners and provide robust performance and accurate predictions [31].…”

Telecom customer churn prediction in context of composite model

“…By establishing discriminant models and regression models to differentiate and classify the customers of the three major operators in China, Zhang and his colleagues can predict the trend of customer churn relatively accurately and can formulate corresponding plans to reduce churn [4]. Soban et al predicted a mixed model of 50,000 entities and achieved a model with an accuracy of 98% under the curve (AUC) [5]. Compared with other methods, the goodness of fit of the model has been significantly improved after the addition of the XGBoost classifier.…”