Churn Classification: An Exploration of Features to Improve the Performance

This work explores the domain expert’s knowledge-based feature engineering for the churn problem. We employ 10-fold cross-validation for parameter tunning and leave-one-out validtion on baselines classifiers. An improvement of up to 9.2% was achieved in terms of the true positive average rate compared to the original dataset, using baselines classifiers. We consider the true positive rate one of the main indicators to measure the churn problem jointly with accuracy and F1. The proposed criteria outperformed the original features between 1.29% and 4.63% in average accuracy using leave-one-out cross-validation. It was also observed that decision trees and ensembles methods performed better than stand-alone classifiers in this problem.