Akademik Veri Yönetim Sistemi
Journal of Industrial Engineering, cilt.32, sa.3, ss.496-512, 2021 (Hakemli Dergi)
The purpose of this study is to provide a descriptive analysis of the assessment of machine learning algorithms to an effective customer churn prediction (CCP) methodology. In the rapidly developing field of Customer Relation Management (CRM), to propose a convenient CCP methodology for retaining the customers who tend to churn, a set of data-mining analyses has been conducted to predict customer churn from a bulky dataset from customers with specific attributes in a telecommunication company by using machine learning (ML) algorithms built in an open-source data mining software, WEKA. Throughout the study, a set of experimental analyses regarding customer churn prediction are conducted by using residential, corporate, and combined datasets with the number of incidences of 195712, 32905, and 228617 respectively a private telecommunication company in Turkey. Six data mining algorithms have been evaluated to predict the customer churn status: Logistic Regression, Naive Bayes, J48, and ELM schemes such as RandomForest, Bagging, and Boosting. RandomForest uses RandomTree, whereas Bagging uses J48 as a base learner. The experimental analyses are conducted with real-world datasets acquired from the company's historical database to validate some decision trees' effectiveness and ensemble ML classifiers to determine the likelihood of future churning customers based on such data mining analyses implemented for CCP. The results show that the J48 outperforms Naïve Bayes based on all datasets, and it provides very similar results as the Logistic Regression classifier scheme. Besides, since Bagging has not solved the large-sized database and J48 has given similar accurate results in the residential and complete data sets, the J48 decision tree classifier can be chosen and Bagging for customer churn prediction.