Customer churn analysis based on machine learning by using data mining techniques in telecommunication sector

In today's increasingly competitive environment, it is necessary to follow the needs, demands, and expectations of customers closely for the enterprises and to respond in the most appropriate and fastest way. It aims to gain customer loyalty by developing mutual relations with customers and thus to provide long term benefit to the enterprise. Today, the cost of earning new customers is much more than the cost of keeping existing customers. Providing promotions, rebates, gifts or benefits to the customers who are anticipated to churn will be able to hinder the churn customer and thus make more profit in the long term. However, if the wrong prediction is made, this causes unnecessary promotions or gifts to the customer. So for the company, this means a waste of unnecessary money. Therefore, it is important for companies to correctly estimate the churn customer. With the help of technology, enterprises can analyze the data they collect from different sources by using various data mining methods and obtain more valid information about the customers and thus develop more effective communication with customers and ensure their continuity. In this thesis, various data mining techniques and classification algorithms of machine learning were used in order to predict the churn on customer data belonging to the telecommunication company. In the data set, there are 7166 customers' data and there is a flag whether the customer churn or not. Also, 328 customers of the data set have churn label. It aims to estimate churn customers with the highest rate. With the train test split, the data set is divided into 70% - 30% training and test data set. Scale and log transformations are performed on data. The 100 most effective features were selected. The performance of the models obtained by classification algorithms is examined. In this study, customers' data are analyzed with machine learning algorithms by using the Logistic Regression, K Nearest Neighbour (KNN), Naive Bayes, Random Forest, Decision Tree, Support Vector Machine and Gradient Boosting algorithms. K Nearest Neighbour and Random Forest have 0.72 accuracy score that the highest accuracy score among algorithms used. Logistic Regression has 0.65 accuracy score, Support Vector Machine has 0.62 accuracy score, Gradient Boosting has 0.61 accuracy score, Naive Bayes has 0.57 accuracy score and Decision Tree has 0.56 accuracy score.