نقش تبیینی با استفاده از خوشه بندی مبتنی بر رفتار در شبکه های مخابراتی

Understanding the individual behavior has shown to be of paramount importance to the triumph of the telecommunication operators to retain customers, enhance their purchasing capacity, and predict the churn rate. Different behavior patterns can be observed for different groups of users. Hence, there is an interesting problem posted in telecommunication network that how to define the users’ role according to their behavior patterns. Traditionally, user behavior characterization methods generally based on their call detail record (CDR), which are user’s individual features, are not appropriate to identify the role in network. In this paper, we develop a new methodology for identifying users’ role based on their behaviors in telecommunication network using the social features instead of their individual features. Experiments have tested on synthetic data and large real datasets, and reveal good results on both of them. Finally, the methodology is not only limited to call graphs but also apply to other networks for role defining.

Many datasets can be described in the form of graphs or networks where nodes in the graph represent entities and edges in the graph represent relationships between pairs of entities. A social network is a graph that the nodes generally represent individuals or organizations and the edges represent relationships among individuals (Wu, Wang, & Zhu, 2008). There is an increasing interest recently in developing algorithms and software tools for exploratory and interactive analysis of social networks, especially on online network and telecommunication network (Anagnostopoulos et al., 2008 and Crandall et al., 2008). Social network analysis methods try to achieve different goals. A well-known issue is the computation of the importance or centrality of vertices. For instance, Google’s Pagerank defines the importance of Web-pages relative to an user’s query and thereby facilitates seeing the most important results first. Other examples are popularity-rankings of users in social networking sites. A different goal in network analysis is the computation of densely connected groups of vertices, which is defined as community discovery. The hope is that these clusters correspond to natural divisions of the network into similar partitions. A third issue in social network analysis aims at computing groups of vertices that occupy the same structural position or play the same role in the network. In the recent years, influence ranking and community discovery have received much interesting, while the notion of role assignment is not as well-known as them. Role assignment try to identify classed of vertices that occupy the same social position, play the same role, or have the same function in the network. It is different from the problem of community partition, for instance, densely connected employees may occupy different positions (like, e.g. manage or secretary). There are not much research work focusing on role analysis in these years. Therefore, according to Scott (2000), social role is the core element of social network analysis. Thus, in this study, we focus on the problem that is to analyze the customer role in telecommunication network based on user behavior pattern. The comprehension of peculiarities of the user behavioral patterns in telecommunication systems are of significant importance in the sense they can give useful information for the operator to better understand their customers. Partition the different customers into various roles enable the operators to build strategies according to different behavior features. The present partition methods might only considering to the customers’ CDR data (Lin, 2007 and Yan et al., 2005) which are mainly standing for individual features. In contrast to these static records, the social features carry information not only about the individual users but also implicitly about their neighbors. These aggregated information are stronger and outweighs CDR records, thus could define better user behavioral pattern. To achieve different goals that specialists in their different designing tasks, user behavior models have been extensively studied and are able to capture particularities from a group of users with a single behavior to multiple classes of users (Maia, Almeida, & Almeida, 2008), including Web, media, E-business, etc. There are also several research works concentrated on the telecommunication field (Dasgupta et al., 2008, Du et al., 2009 and González et al., 2008), while few targeted on partition different kinds of customer roles. The characteristic of our approach which different from the other behavior-based methods in that it combines both of the node identity and the relational data. By taking into account of the network structure, the role defining process can benefit from two main respects: firstly, it takes account of the interaction data of the nodes which is very useful for understanding the node’s value; secondly, it can give a global view instead of a static view of the data. The contributions of this paper include the following: providing a proper method, that is bringing in the structure information to the telecommunication network analysis; observed from the user behavior patterns form by the social features, we can partition the users into different roles according to which operators can make different strategies; we can also apply the methodology used in this paper for user role identification to social networks in general. The remaining of this paper is organized as follows: Section 2 presents the related work. The data collection and feature sets are described in Section 3. Section 4 presents the role analysis process. Section 5 describes the experiment results and Section 6 concludes this work and expresses future directions.

In recent years, we witness dramatic increase in the competition among telecommunication companies in order to detain their current subscribers and acquire new ones. Identifying user role within an environment in which we know nothing a priori is a challenging but practical task. For instance, identify the user role in the other operators according to their calling behavior to our local subscribers. In our study, we choose the social features and use K-means clustering algorithm, along with Mahalanobis distance as distance measure and BIC as our stopping criteria. Therefore, the number of clusters, and the ultimately user behaviors discussed in our study, differs from the nature of the network and the data available. Finally, the methodology we discussed in this study is not limited to telecommunication network, but could apply to many other networks. For future directions, we could investigate to improve the performance of our methodology and how to apply our role analysis method to the practical application, for instance, how do the users with different roles related to the churn rate.