Article Preview
Top1. Introduction
During the last decade competition became a real concern for telecommunication providers (Bin et al., 2007). Thus, operators are poised to find new methods to enhance the quality of their services and diversify periodically their portfolio to retain the existing customers and attract new ones. While the primary focus of each telecom service provider is to provide customer service satisfaction, preventing subscribers from churning remains a huge challenge. Churn in telecoms is the term used to collectively describe the ceasing of customer subscriptions to a service (Huang et al., 2010), and if one customer cancels his/her service and switches to another operator, this customer is considered as a churner. Converging lines of evidence showed that the cost for customer acquisition is much greater than the cost of customer retention (in some cases~20 times more expensive) (Vafeidis et al., 2015). Thus, it is compulsory to telecom service providers to identify unsatisfied subscribers to prevent them from churning. For this, developing reliable predictive models to predict customer churn are crucial for the business management of the telecom industry.
Telecommunication companies consider customer churn a real and serious common business problems that should be addressed very carefully to avoid the loss of potential subscribers. In our work, we focused on the prepaid subscribers, a category of customers who can terminate their service subscriptions and switch to another telecom provider without prior notice. For instance, we found that, in one major telecom operator in Morocco, the churn rate of prepaid subscribers is significantly higher than postpaid subscribers (Figure 1). While it is possible that such high churn rate in prepaid subscribers may be is due to factors related to high cost offer, low quality of the service, and/or high customer service dissatisfaction, being able to analyze and monitor customer behavior in time gives companies the opportunity to execute preventive measures for retaining them.
Figure 1. Churn rate comparison during 12 months
Several algorithms have been developed (Vafeidis et al., 2015), however it is not clear which of those models better fit for detecting the churning customers. To evaluate better the churning impact on customer’s network, several experiments have been conducted using the state-of-the-art machine learning techniques with special emphasis on deep learning algorithms and convolution neural networks.
ConvNets (CNN) have been proved to have a very good performance in different area of research, including images and video recognitions (Simonyan et al., 2014; Yang et al., 2015; Hatami et al., 2018), natural language processing (Zhang et al., 2015), and speech recognition (Noda et al., 2015) by extracting high-level features from a large set of data. In addition, CNN has been demonstrated to have a very good performance in image processing tasks (Szegedy et al., 2015; Russakovsky et al., 2015). Thus, it would be of interest to use this technique to predict the customer churn by analysing images of represented customer’s behavior.