Hybrid Features for Churn Prediction in Mobile Telecom Networks with Data Constraints

Abstract

In a competitive Mobile telecommunications market, the customers want competitive pricing and high quality of service. A customer won't hesitate to change their telecom service provider if he/she does not find what they are looking for. This phenomenon is called churning. The telecom service providers often find that the cost of acquiring a new customer is much more that the cost of retaining one. Hence telecom operators are focusing their marketing strategies toward targeted customer retention campaigns and this is known as Churn management. One of the primary tasks of Churn management is to build an effective churn prediction models that can predict customers who are most likely to churn. The primary idea is to create profile of a customer using various data sources including call patterns, contractual information, billing, payment, customer service calls, demographic profiles and then predict the probability that he/she will churn based on his/her features. The apparent drawback of these approaches is that they require access to numerous other sources of information apart from Call Data Records (CDRs). More importantly, these models do not take into account any social influence. In our present work, we recognize the importance of the role played by social ties understanding the causal behavior of customers, and incorporate a novel feature of the social aspects of customers' social group along with the traditional individual customer profiles with potential practical implications. We propose hybrid feature sets that are based not only on the features extracted from CDRs but also on the changes in these feature sets combined with the changes in the social group patterns that would give improved performance over existing models with similar data constraints. Despite the data constraints, we demonstrate through our experiments that our model achieves improved prediction performance using these hybrid feature sets extracted from the CDRs as well as mobile social graphs even with our data constraints.