Most of telecommunication companies consider the customer as the most important asset for them. For that reason, nowadays, a challenging problem that encounters telecommunication companies is when the customer leaves the company to another service provider for a reason or another . In most cases, this churn can happen in rates which seriously affect the profitability of the companies since it is easy for the customers to switch companies.
In market, where the competition between the telecommunication companies grows rapidly, companies have shifted their focus from acquiring new customers to retain their existing ones [1–3]. Basically, churn is one of these significant problems and companies started to seek new Business Intelligence (BI) applications that predict churn customers. When the company is aware of the percentage of customers who leave for another company in a given time period, it wouldbe easier to come up with a detailed analysis of the causesfor the churn rate and understand the behavior of customersthat unsubscribe and move to other business competitor. Thishelps in planning effective customer retention strategies for that company .
Among many approaches developed in the literature for predicting customer churn, supervised Machine Learning (ML) techniques are the most widely investigated [5–9]. Supervised ML concerns the developing of models whichcan learn from labeled data. ML includes a wide rangeof algorithms such as Decision trees, k-nearest neighbors,Linear regression, Naive Bayes, Neural networks, Supportvector machines (SVM), Genetic Programming and many others.
For example, in  authors conducted a comparative analysis of linear regression and two machine learning techniques; neural networks and decision trees for predicting customer churn based on variables related to customer complaints. Their dataset was balanced with 50% ratio of churners and non-churners. Authors showed that neural networks have good capability for predicting churners. In another work, authors in  investigated the use of Genetic Algorithms based neural network models to help in predicting churn in cellular wireless services. Their models show better accuracy compared to statistical z-score model.
In general, most of authors in the literature focused on using ML techniques for churn prediction but few of them focused on the importance of churn factors in the prediction. Identifying important factors can greatly support customer relationship management in telecommunication companies to plan effective customer retain strategies. For that reason, in this research, we use a multilayer perceptron neural network not for predicting customer churn but also to give an insight on the relative importance of each input variable regarding customer churn. We investigate two different approaches based on neural networks for identifying important variables. The first is based on error change and the second is based on weights contribution in the...