Journal of Industrial Engineering, University of Tehran, Special Issue, 2011, PP. 25-37 25
Customer Churn Prediction Using Local Linear Model Tree for Iranian Telecommunication Companies
Mehdi Fasanghari 1 and Abbas Keramati 1
Department of Industrial Engineering, College of Engineering, University of Tehran, Tehran, Iran
(Received 16 November 2010, Accepted 11 July 2011)
For winning in global competition, companies need to recognition and monitoring of customer’s behavior to forecast their behavior and desires earlier than competitors. This research tries to recognize the attributes which lead to customer churn. For this, behavior of 3150 subscribers of an Iranian mobile operator, has observed during one year and trends of them has analyzed by a customized LLNF model. For this purpose, the application of the locally linear model tree (LOLIMOT) algorithm, which integrates the advantage of neural networks, tree model and fuzzy modeling, was experimented.
Results suggest that dissatisfaction of customer, his/her usage from services and demographic attributes have significant effect on decision to churn or retention. Furthermore, the active or inactive subscriber situation has mediation effect on his/her retention.
Keywords: LLNF, Customer churn, LOLIMOT, Fuzzy Logic, Neural network, Prediction,
Mobile service provider
In age of communication, it is natural that telecom industry possesses one of the highest growth indicators among other industries. Besides, among different telecom industries, one of the fastest rates of growth belongs to the mobile service industry that its contribution in everyday communication of people is extraordinary increasing and growing on the verge of overtaking fixed phone. This growth speed is not only owes innovation in Mobile Communication technology, but also owes intense competition and ruthless operators under the minimum regulations governing the market of this industry. Mobile service market along with the growth of its industry shows the remarkable growth that cannot be due only to increase in the number of subscribers and the increasing diversity of new services and ignore competitive arena of them. This surprising and integrated development of industry and market caused the operators despite being a young industry, trying to get new customers and start their marketing blitz gradually and attract customers of other operators [1, 2]. Obtaining new customers is much harder and more expensive than keeping existed customers, and according to the study has been done in 2004, the cost is equivalent to 300 dollars versus 20 dollars in 1995 and 300 dollars versus 25 dollars in 2004 . Part of this fact is due to the suppliers that have significant and valuable information about their customers and analyze it to understand their behavior and preferences. Additionally, in a developed market, attracting new customers needs to ward off them from other operators and having a more attractive stimulus.
Keeping old customers, especially inservice markets such as mobile cell, except the cost of attracting customers, have the opportunity value for mobile operators. It means that the provider can provide additional and new services to customer and earn more money. For this reason, the loss of existing customers not only reduces revenues and imposes costs of attracting new customers, but also leads to the loss of potential revenues.
* Corresponding author: Tel: +98- 21- 82081067 Fax: +98- 21- 88013102 Email: [email protected]
Therefore, with the growth of telecom companies in the global market and maturation of these services, customer churns management has become a fundamental concern of these companies. Currently, the churn rate of customers in such companies in Europe, North America and East Asia has been reported in a range of 20% to 40% [1, 2, 4]. This high rate of customer churn is in an industry that its cost to obtain a new customer is five to 10 times more than maintaining an existing one . Calculated values in the mobile service market in America are equivalent to 300 dollars versus 25 dollars .
Living in the global village and taking the path of fast globalization has made communication development, and especially mobile services, for developing countries such as Iran necessary. Providing mobile services in Iran began in 1991 by Telecommunication Company of Iran as the first operator. Telecommunication Company was a 12-year monopolized in this industry, until entering Taliya Company as the first private mobile service provider in June 2003 with the first series of MMC card, and ended the monopolized market. Irancell Company started its services as the second operator of mobile services in Iran, with the participation of MTN Company of South Africa (the largest mobile operator in Africa and the Middle East now). So far, 15 foreign companies from Canada, Great Britain, Russia, Malaysia, China and the UAE stated that they would have been ready to participate in the development of the third mobile operator, according to the program to cover 17 million users. Now considering a competition in similar global markets, as well as population growth and the growing needs of mobile communication, strong competition among operators to increase their penetration and also increasing their contribution in the Iranian market can be predicted.
Imparting the general policies of Article 44 of the constitution which quickly accelerated privatization in Iran, Iranian mobile service market will experience stronger competition, particularly in the field of MMC card. The first operator is on the verge of privatization, and after that management ability, true policies and timely marketing make survival of this great organization possible in the absence of strong government institutions.
In this strong competition, the winner will be someone who protects her customers as valuable assets while trying to attract new customers, and use scientific methods and useful tools in this way.
Considering challenges and thriving future of this market, a case study research on one of the active companies in mobile services in Iran has been done, with the help of survival analysis tools to understand features that lead to customer churn.
Although according to some researchers, studies carried out in customer churn have not been focused and organized so far , but they can generally be divided in two categories:
Articles and researches had been done to show what data mining can do in customer relationship management (which henceforth will be mentioned briefly by CRM) and provide reasons to prove fitness of data mining tools to analyze CRM processes and also articles proposed in order to introduce specific data mining methods for solving problems facing the CRM in the organization.
Field researches that analysis customer churn with respect to customer behavior in different organizations.
A. Importance of data mining in Customer Relationship Management
In this article, CRM is defined as a multiple layer concept that data mining is one of its layers. In his study, considering the results of data mining analysis about customers is required for an efficient CRM system. Also two examples of data mining techniques were introduced and their applications in two case studies of CRM were described accurately. Valletti and Cave, in 1998, analyzed expanded competition in the mobile industry from the perspective of different operators’ strategies . Fullerton also followed the same goal in his research about the telecommunications market in America, studying the relationship between market structure and market performance and has been shown that how understanding the market structure can help predicting market performance and reduce the customer churn to its minimum.
B. Analysing the behaviour of customers and their churn
The second category includes those researches that analyze issues related to customer churn in mobile phone services and have been highly regarded in recent years. Song and Kim, in 2001, evaluated the effect of change in mobile market structure of the Korean on customer churn, utilizing the simulation approach . Choi, Lee and Chung (2001) also examined the effect of business strategies of five major mobile service provider companies in Korea’s on customer loyalty . Kim and Kwon in 2003 began to study factors that customers consider when choosing their mobile operator . The research results indicate the effect of discount of intra network calls and the quality of communication on selecting the operator. In 2004, Kim and Ion trace on 973 users of five major mobile service operators in South Korea and identified subscriber churn and loyalty features . Their study has shown that the probability of modifying operator by a customer depends on a satisfaction level related to his operator service features, including conversation quality, tariff levels, the device provided by operator and brand reputation. Factors such as conversation quality, type of device offered by the operator and brand reputation also impress the customer loyalty, and this loyalty is measured by a customer intend to recommend the operator to the others. Not existing significant relationship between subscription duration and customer loyalty, indicated “lock-in” effect among customers that they can be named as a sham loyal, those who don’t end their subscription only because of avoiding additional costs.
The survey research of Gerpott, Rams and Schindler, in 2001, hold on the mobile market in Germany showed that maintaining customer, his loyalty and satisfaction are correlated .
In their study, structural differences in three concepts of customer maintenance, customer satisfaction and customer loyalty and also their correlation had been investigated. Their analysis is based on a two-stage model in which overall customer satisfaction significantly affects the loyalty, and the loyalty affects his intent to leave or stay in touch with the operator. Price of mobile services, perception of benefit from services and portability of numbers between different operators were considered as variables related to the provider that has a lasting impact on customer churn.
In 2006, the telecommunication industry’s research was published in South Korea . This study, entitled ” customer churn Analysis: churn factors and the effect of transition stage churn on Korean Mobile Communication Industry” by Ahn, Hana, and
Lee that studied churn factors based on data related to transactions and payment of subscribers and claimed that cost of modification and service consuming also affect the decision of leaving or sustain. Most of the previous studies focus on the direct impact of independent variables on customer churn while this study states that status of customer is as intermediaries between the customers churn signs and complete churn. Another study, by Seo, Ranganathan, and Babad was published in 2008 that focuses on understanding the factors that led to the customer churn . To understand these issues two basic questions come in mind: 1- How the customer satisfaction and cost of modification factors, such as subscription time, complexity of services and quality of communication, help customers sustain and 2- How a customer demographic characteristics such as age and sex affect their choices and lead to differences in their approach of leave or sustain.
Customer churn analysis in the mobile market of Iran is a time consuming and complex problem which has many criteria for consideration. According to the literature review, the statistical analysis methods are useful for statistical analysis of the customer churn data. Thus, it is recommended in this paper to propose a customized neural network to come across the customer churn complexity in the mobile market of Iran as a new contribution.
One of the restrictions in this paper is the limitation for access to all the customer churn parameters in the mobile market of Iran. Most of the parameters are not available in this industry, and some of the parameters are not clear enough for considering in customer churn analysis. Therefore, it is necessary to constitute the customer churn criteria framework in the mobile market of Iran according to the available information.
Customer churn prediction in Iranian mobile operators company is done in this paper according to the proposed methodology in Figure 1. As presented in this methodology, the data should be selected from a database which is belonged to a mobile operators company in Iran, and the dataset should be large enough and have the mentioned criteria of the proposed customer churn framework as the attribute of each customer data in dataset.
The dataset was picked up from the database of a mobile service provider in Iran. According to the variables which can be extracted from the selected dataset, and as for the developed customer churn framework in this research, randomly, 5000 records are extracted from the available dataset, in which 3150 records were prepared for use in this research as the other records was not complete in all the attributes assigned for each record.
Figure 1: The proposed research methodology
We customized a local linear neural network for modeling the customer churn prediction with LOLIMOT learning method. 80% (2520 records) of data was assigned for the training process of the developed neural network, and the rest (630 records) was assigned to the test process. Then, for evaluating the developed method, the multilayer perceptron was selected for comparison. Finally, the obtained results were comprised with the most popular evaluation criteria (RMSE and R2).
C. Analysis of Customer Churn Framework
The usefulness of a theory depends on appropriate replication, development, and generalization to create new insights and add it to existing knowledge collections . Reviewing the literature and existing models in customer churn, the model that was proposed by Ahn and her colleagues (2006) was the most comprehensive model that has been studied and was chosen as a foundation for developing the framework of this study because in addition to involving all factors that had been studied in other researches , the effect of some new churn factors is also examined. In addition to measuring the direct effect of these factors on customer churn, the effect of customer status is also investigated and illustrated in Figure 2.