Identifying Key Learning Algorithm Parameter of Forward Feature Selection to Integrate with Ensemble Learning for Customer Churn Prediction

Abstract

The Telecommunication has been facing fierce growth of customer data and competition in the market for a couple of decades. Due to this situation, an analytical strategy of proactive anticipation about customer churn and their profitable retention is inevitable for Telecommunication companies. To nip this problem in the bud, a lot of research work has been conducted in the past, but still the previously introduced churn prediction models possess their own limitations, such as high dimensional data with poor information and class imbalance, which turn into barriers while being implicated in real life to attain accurate and improved predictions. This study has been conducted, basically, to identify the key Learning Algorithm parameter of Forward Feature Selection (FFS) for dimensionality reduction which can be further integrated with class Imbalance Handling Technique and Ensemble Learning (EL) to attain improved accuracy. The core objective of this study is to turn an imbalanced dataset into a balanced one for Ensemble Learning (EL) Model of Customer Churn Prediction (CCP). This study concluded that Logistic Regression (LR) based Forward Feature Selection (FFS) can outperform with Oversampling Class Imbalance Handling Techniques and Ensemble Learning (EL) by scoring 0.96% accuracy, which is the highest accuracy against benchmark studies. The resulting methodology has been named as the Logistic Regression Learning based Forward Feature Selection for ensemble Learning (LRLFFSEL) and applied over Orange dataset with 20 features and 3333 instances. In future this methodology can be evaluated over a bigger dataset and combined with some data optimization techniques to improve its accuracy.