Resource type
Thesis type
(Thesis) M.Sc.
Date created
2021-06-03
Authors/Contributors
Author: Ren, Yingwen
Abstract
With the rapid growth in the telecom market, there is an emerging trend to focus on customer retention, which is a critical factor for designing future customer incentive strategies to help a company manage customer relationships. Our main contribution is to build an effective churn prediction system for a telecom company providing real-time communication to predict whether a customer may cease to do business with the company, i.e., stop using the service provided by the company to make phone calls. Due to the dynamic market environment, developing such a system is challenging, as it should not involve frequent retraining processes, leading to a high computational cost. Many different techniques are available to identify customers who are most likely to leave, however, which technique is the most suitable and applicable in practice is not clear because the performance of prediction methods depends heavily on the characteristics of the data. In our thesis, we implemented and evaluated two methods, namely MLP (Multilayer Perceptron) and WTTE-RNN (Weibull Time To Event Recurrent Neural Network), and the model evaluation is based on accuracy and computational cost. We conducted experiments on the real-world dataset containing customer call activity records, experimental results demonstrate that the model performance of MLP is better than the WTTE-RNN, achieving a higher AUC, precision, and Recall. Considering the computational cost, the WTTE-RNN takes more time than the MLP as the WTTE-RNN needs to be retrained, it cannot be directly applied for new data. Furthermore, a detailed feature engineering process was presented in our project, especially how to extract temporal call behavior from raw data. A user-friendly interface was implemented, in order to let users better use our churn prediction system.
Document
Identifier
etd21429
Copyright statement
Copyright is held by the author(s).
Supervisor or Senior Supervisor
Thesis advisor: Wang, Ke
Language
English
Member of collection
Download file | Size |
---|---|
input_data\21650\etd21429.pdf | 2.83 MB |