Using PCA to Predict Customer Churn in Telecommunication Dataset

Failure to identify potential churners affects significantly a company revenues and services that can provide. Imbalance distribution of instances between churners and non-churners and the size of customer dataset are the concerns when building a churn prediction model. This paper presents a local PCA classifier approach to avoid these problems by comparing eigenvalues of the best principal component. The experiments were carried out on a large real-world Telecommunication dataset and assessed on a churn prediction task. The experimental results showed that local PCA classifier generally outperformed Naive Bayes, Logistic regression, SVM and Decision Tree C4.5 in terms of true churn rate.