Customer churn predictive analysis by component minimization using machine learning

Abstract

Telecom industry faces crucial competition as new deals are introduced in market daily. Many operators are finding difficult to identify the potential customers with the drastic variations in offers of their competitors. Hence reducing the customer churn is the biggest challenge for telecom operators since the reason for churn is unknown. Due to volume of data, they are not able to predict the cause of customer churn. Appropriate machine learning algorithms help to understand why subscribers leave by finding the relationships between data. This paper applies classification algorithms to predict the behavior of customer retention on a telecom dataset extracted from Kaggle. The performance of the dataset after dimensionality reduction using PCA is also assessed. A comparative analysis on different classification algorithms are made based on the performance metric such as accuracy, precision, recall, log loss and f-score. The developed model performance is shown using ROC and AUC curves. Experimental results shows that after applying PCA, the kernel SVM is found to be effective with the accuracy of 95.5% compared to other classifiers.