Customer churn analysis in telecommunication industry

Consumption preferences of people vary depending on their needs. And, institutions investing in clients cannot predict these preferences. Especially, customer-oriented institutions try to gain new customers and prevent customer churn by satisfying existing customers. Telecommunications industry is one of the customer-oriented industries. Telecommunication companies also want to gain customers, without losing existing customers. At this point, they engage in prediction of customer churn using various methods. In this thesis study, customer churn analysis was performed with classification algorithms, which are among the data mining and machine learning methods. In carrying out this analysis, the Cross Industry Standard Process for Data Mining (CRISP) model, which is one of the machine learning process steps, was used. The thesis was explained through the steps of the CRISP model from identification of problem to model selection. The performances of the models obtained by the classification algorithms were evaluated by the cross-validation and hold-out performance methods. The 4-fold, 5-fold and 10-fold cross-validations were used. Models built with decision tree algorithms in performance evaluation with 4-fold, 5-fold and 10-fold cross-validation showed better performance than the other models. The performance of the best performing C4.5 decision tree was approximately 0.98. The C4.5 decision tree was followed by the models created with ID3, Gini decision trees, k-nearest neighbors and Bayes algorithms, respectively. Although the k-nearest neighbor algorithm comes after the decision trees, its performance was closer to that of C4.5 decision tree. In the performance evaluations performed on the training-test dataset with the 60-40%, 75-25% and 80-20% separation ratios with the hold-out method, respectively, the best-performing was the C4.5 decision tree, similar to that of k-fold cross-validation performance. This was followed by ID3 and Gini decision tree and k-nearest neighbor algorithm, with close values as in k-fold cross-validation performance method. The Bayes algorithm had the worst performance. Since the k-nearest neighbor algorithm ID3 and Gini perform better at random distinction with hold-out of decision trees. A study on data visualization has also been carried out through R which is used as a data mining program. In addition to these studies, C4.5, which gives the best result from the classification algorithms, has been rendered dynamic by making web application with Shiny from the R packets generated by the decision tree algorithm.