Churners Prediction Using Relational Classifier Based on Mining the Social Network Connection Structure

Abstract

Customer churn prediction models aim to indicate the customers with the high tendency to churn, allowing for improved efficiency of customer retention operations and reduced costs associated with the attrite event. This paper proposed a data mining model to predict churn customers using Call Detail Records (CDR) data in the Telcom industry. CDR data are valuable for understanding the social connectivity between customers through call, message or chat graph but do not immediately provide the strength of their relations or present adequate information of how to churn node diffuse it is influenced to all neighbor nodes within the call graph. The main contribution of this paper is to propose a data mining model to predict Potential churners based on social ties strength and churn diffusion through network nodes. The model formulated by extracting the predefined number of social network communities that denoted the fundamental constructions for figuring out the connections structure of the call graph using Non-negative matrix factorization approach. Then, for each community quantifies the social ties strength using Node interaction and similarity algorithm. Then, incorporate social strength ties in an influence propagation model using are receiver-centric algorithm with Logistic Regression classifier to predict the set of the customers at risk of churn. Experiments conducted on Telecom dataset and adopting of AUC, ER, Accuracy, F-score, Lift, ROC, KS and H measure as evaluator metrics show that using of influence node diffusion based on ties strength boosted the performance of the proposed churn model between 84% to 91% in terms of the accuracy metric.