Performance comparison of six Data mining models for soft churn customer prediction in Telecom

Abstract

Due to a high competition in the market, the telecom operators are affected by churn, therefore it is very important for them to identify which users are likely to leave them and switch to the competition telecom company. This research uses data on behaviour of the users from telecom systems that serve to identify patterns in behaviours and thereby recognize the churn. Creating new definition of prepaid soft churn based on multiple conditions is valuable contribution of this paper. At preparing data, a selection of useful attributes was made using the Principal Component Analysis (PCA). The normalization of the attribute values has also been made in order to obtain a proper balance of the influence of all the attributes. Common problem with telecom churn prediction data is imbalance, taking into account the target variable. Such a case is also in the data used in this paper, where the percentage of churners is 12%. Comparison of undersampling and oversampling was performed as a method for resolving the data imbalance problem. Data sets with undersampling and oversasmpling have been used to train the decision tree, logistic regression and neural network algorithms and therefore six prediction models for detecting the churn of the Prepaid users in the telecom were created in this paper. Performance analysis and comparison of the six developed Data mining models was also performed.