Thesis Type: Postgraduate
Institution Of The Thesis: Middle East Technical University, Faculty of Engineering, Department of Computer Engineering, Turkey
Approval Date: 2015
Thesis Language: English
Student: Pelin Ercan
Supervisor: FERDA NUR ALPASLANAbstract:
Nowadays, more and more companies start to focus on Customer Relationship Management (CRM) to prevent customer loss. The early detection of future churners is one of the CRM strategies. Since the cost of acquiring new customers is much higher than the cost of retaining the existing ones, it is important to keep existing customers. Churn is an important problem for the game companies as churners impact negatively for potential and existing customers. Data mining can support an individualized and optimized customer management to avoid customer loss. In this thesis, the problem of player churn in Pishti Plus, which is a multi-player social game, is studied. The purpose is the detection of churners by using the first 24 hours log data of the players. Data used in the prediction model is selected using correlation based filter method. Results of Bayesian Network, Logistic Regression, Sequential Minimal Optimization (SMO), and Simple Classification and Regression Tree (CART) algorithms are compared and an early prediction model is built. Ensemble methods are applied to improve the accuracy of the model. Results indicate that Simple CART algorithm is more successful for predicting churners. The built model predicts the churners with 68.20 % accuracy.