Research On Telecom Customer Segmentation Based On Semi-Supervised Affinity Propagation Clustering

In recent years, with rapid development of the telecommunications market, thecustomers of telecommunications show characteristics of segmentation and diversity, and thecustomer is the key to the development of fundamental and corporate survival.Face to the vast amounts of customer data and information, how to derive usefulinformation for operators to maintain the companies and develop customer decision-makingsupport, has become the primary issue of the mobile communication operators. Only by usingthe effective methods and tools to depth analysis of vast amounts of customer data, candevelop a precise marketing plan, and get a higher return on investment.Customer segmentation is one of the core functions of customer relationshipmanagement system, available to customers, and provides comprehensive customer supportand maintain the value-added.In this paper, we analyze and discusse the customer segmentation results too general,segmentation variables and descriptive variables irrational and other related issues in depth.The research work is concentrated by the following aspects:(1) To determine the customer segmentation variables and their description based on themobile industry massive customer data resources, so that better results can be divided intosegments of solving the problem of too general breakdown of the results.(2) To propose a density-sensitive semi-supervised AP clustering algorithm, byintroducing semi-supervised paired and a density-sensitive distance constraints introducedinto the AP clustering. Compared to conventional AP algorithm, the proposed method hasbetter clustering performance, especially for non-convex data set clustering effect to havebeen significantly improved.(3) Based on the density-sensitive semi-supervised affinity propagation clusteringalgorithm, combined with the appropriate customer segmentation theory to construct atelecommunications customer segmentation system model.Experimental results show that the proposed algorithm on the clustering performanceimproved significantly. In practical applications, customers can be more clear on thebreakdown, thus providing a more scientific decision-making guidance for operators.