Article Preview
TopIntroduction
Due to the increased competition in the telecommunication industry, customer churn has become a serious issue for major telecommunication companies. Customer churn is the process of customers switching from one service provider to another anonymously (Umayaparvathi & Iyakutti, 2012). Finding a solution for customer churn means that companies should be able to predict in advance customers who are more likely to leave. Retaining customers has become a strategic point to protect revenues for telecom sector investors. To address this situation, telecom operators are forced to take required retention policies and spend more amount of investment on retaining existing customers as opposed to acquiring new ones (Idris & Khan, 2012).
Unlike terabyte databases, which typically contain images or multimedia streams, telecommunication databases mainly contain numerous small records describing transactions and network status events (Koutsofios et al., 1999). These facts make the analysis of such massive data a big challenge that consumes a lot of time and needs considerable effort and expertise. The main obstacle to dealing with such a situation is the time necessary to do the treatment and deduct the potential churners; the allowed time should not exceed a few days, to anticipate the departure of the client and give the marketing department time to offer alternatives likely to retain customers. It is important to note that the investment for customer retention is very critical, and errors in detecting real potential churners can cost companies considerable amounts.
Customer churn prediction is mainly a classification problem. The objective then is to predict churners before they leave the company. We consider binary the labels target, Churn, and Not Churn for every customer. It is important to note that the proportions are not equivalent and that a minimum percentage is likely to churn if we consider a random sample of customers. This phenomenon is called unbalanced datasets. To process those customers’ data and arrive at a realistic model, it is often important to consider a large number of attributes to give visibility on all the factors that can cause a customer to churn.
The approaches and criteria for choosing the best model are very diverse. Hargreaves (2019) employed the Logistic Regression model, as it requires little running time compared to other complicated machine learning algorithms and its output is also easy to interpret. Induja and Eswaramurthy (2016) put into use the kernelized extreme learning machine (KELM) algorithm, proposed to categorize customer churn patterns in the telecom industry. The primary strategy of the proposed work is to organize the data from the telecommunication mobile customer dataset. The data preparation is conducted by using preprocessing with the expectation-maximization (EM) clustering algorithm.
Other than studying a combination of methods and algorithms, Kamalakannan (2018) used a normalized k means algorithm for dataset preprocessing and then selected the attributes from a preprocessed image using the minimum redundancy and maximum relevance (mRMR) approach. It tends to select attributes with a high correlation with the class (output) and allow correlation between themselves. The prediction is examined with the help of a support vector machine with particle swarm optimization (SVM with PSO). Gunay and Ensari (2019) analyzed well-known machine learning methods, including logistic regression, naïve Bayes, support vector machines, and artificial neural networks, and proposed a new prediction method. Abbasimehr et al. (2014) performed a comparative assessment of four major ensemble methods (bagging, boosting, stacking, and voting) based on four base learners, i.e., decision tree, artificial neural network, support vector machine, and reduce incremental pruning, to produce error reduction (RIPPER). The training process was made using the synthetic minority over-sampling technique (SMOTE). The research shows that ensemble learning techniques brought an important improvement to the predictions.