Research On Telecom CRM Based On Complex Networks

As the competition of telecommunications industry continuously increasing and the information technology rapidly developing, telecom operators have focused on the frontiers such as customer service and market analysis instead of the initial part such as production. Correspondingly, telecom companies have changed their management model from"Information-centric"to"customer- centric". It is not difficult to see that the customer resource occupies a central role in the telecom business operations. To enhance their competitive position in the market, the telecom operators should fully understand the customers and distinguish the differences between them, thus the telecom operators can provide quality services, so as to increase their customer satisfaction and loyalty, and finally make a profit.Customer relationship management (CRM) system was developed with the rise of e-commerce and it has demonstrated vigorous vitality since it's generated in the United States in 90's of the 20 century. Industries pay more and more attentions to CRM system in view of the fact that this system contains efficient information management as well as complete data resources processing technology. Telecom industry is one of the earliest industries who implement the CRM system since their awareness that the system has great advantage in many ways. The research on telecom CRM focused on many different points, the customer segmentation is an important aspect in the research of telecom CRM. It is usually achieved by clustering method in data mining. In the process of implementing customers clustering segmentation in the existing business support systems and CRM systems, telecom industry in China is mainly based on the value or the social characteristics of customers.It's not uncommon that the concept of complex networks, as well as its static characteristics, is used in the field of telecom CRM. The complex network represents an abstract model of a complex system, and its ultimate goal is to describe the characteristics and behavior of complex systems accurately. On the other hand, the customer relationship network in telecom industry, in which vertices represent telecom customers, edges represent the calling relationship between customers, and the weight of edges represent corresponding number of the calls, is clearly a social network. The customer relationship network can effectively reflect the common characteristics of consumer behavior in customer groups with different social backgrounds. Therefore, it's of great value to take the customer relationship network as a complex network, and to apply its static characteristics in the telecom customer segmentation analysis.We did a comprehensive analysis on telecom customer relationship management, data mining, complex networks, and many other technologies. On this basis, we summarized and analyzed the shortcomings of k-means clustering algorithm which is commonly used in traditional telecom customer segmentation. We also proposed a cgk-means clustering algorithm. In this algorithm, we take the attributes of customers'consumer behavior as the segmentation variables, and select the initial value under the guidance of the static characteristics of the customer relationship network. We ran the two different clustering algorithms in the data set for testing, and evaluated the algorithms by the measure of purity. What's more, as the data volume in practical application is often very large, we also improved the cgk-means algorithm.The research in the telecom market shows that the property of customers'social background and their consumer characteristics are closely linked in the telecom industry. And the so-called"social background"here gets the best expression in the customer relationship network. The cgk-means algorithm selects the properties of customers'consumer behavior as segmentation variables in order to maximize the differences among customers. In order to determine the initial cluster centers, the algorithm first calculates the eigenvalue for each node using the two static characteristics of the customer relationship network - centrality and clustering coefficient. Since the calling relationship between customers can effectively reflect the characteristics of consumer behavior, the algorithm then takes the calculated eigenvalue as the primary key to sort the nodes in non-ascending order. Next, algorithm selects the initial cluster centers refer to the threshold in the consumer behavior vector space.The application of cgk-means algorithm in the telecom industry has improved the inherent shortcomings of traditional clustering algorithm. We have got more accurate clustering results, thus the telecom operators are able to provide proper telecom products or services for the customers who share similar characteristics of consumer behavior and demands more effectively. The realization of one to one marketing has great significance for the telecom enterprises to improve their competitiveness and service quality.In addition, the paper used correlation analysis and factor analysis in statistical theory when in practical use, in order to reduce the data dimension and filter out the correlation between the variables.The main features and innovation points of the paper are as follows:1. Before clustering, implement two operations in statistical theory- correlation analysis and factor analysis, in order to remove the explicit and implicit correlations between the variables and to reduce the data dimension.2. Combine the vector space of telecom customers'consumer behavior with the complex network space of telecom customers'calling relationship organically. Implement the operation of telecom customer segmentation to make the result more comprehensive and effective.3. When updating the initial cluster centers, cgk-means clustering algorithm achieves two different methods: real-time update method and batch update method, using a parameter for distinguish.4. When selecting the initial cluster centers in the non-ascending sequence, the cgk-means clustering algorithm selects the nodes which satisfy two conditions at the same time instead of simply selecting them accord to the order. One of the conditions is that the eigenvalue of the node is large enough. The other condition is that all the distances between current node and the selected nodes are longer than a certain threshold. Thus, it can avoid the condition that clustering result is not good, as the selected initial cluster centers are too concentrate.5. Improve the cgk-means clustering algorithm by using common library functions in matlab, improving the efficiency of the algorithm, making it adapt to the situation that the amount of data is large.