A Hybrid Prediction Model for Customer Churn using DBSCAN and Stacking-Based Classifier

Abstract

Customer churn is a key concern in numerous companies, including the telecom industry. Decision-makers and business analysts feel that retaining existing consumers is less expensive than acquiring new clients. To provide a retention solution, customer relationship management (CRM) analysts must recognize customers who intend to quit the company and understand patterns of behavior from existing churn customers' data. This paper provides a Hybrid Prediction Model (HPM) that employs classification and clustering approaches to predict client attrition. To pick the key characteristics, the proposed model uses an RFE algorithm, filter method, Boruta algorithm, and correlation matrix, as well as Density-based Spatial Clustering of Applications with Noise (DBSCAN) to discover and eliminate outlier data. In addition, it uses stacking classifier to categorize consumers, tuning threshold to handle data imbalance, and k-mean algorithm to segment churning customers—whom the stacking classifier has classified—into groups to make group-based retention offers. The Telecom dataset is not publicly available due to protect the private information of customers. Thus, the data used in this study was obtained from Openml Website. The data set contains 5000 observations and 21 variables. Moreover, the proposed model can be easily adapted and applied on larger datasets. Several metrics are used to evaluate the proposed hybrid prediction model, including accuracy, recall, precision, receiving operating characteristics (ROC) area, and f-measure. The results show that our hybrid model outperforms single techniques. Furthermore, when changing the threshold in terms of recall metrics, the model performs better.