Churn Prediction System for Telecom using Filter-Wrapper and Ensemble Classification

Churn prediction in telecom is a challenging data mining task for retaining customers, especially, when we have imbalanced class distribution, high dimensionality and large number of samples in training set. To cope with this challenging task of churn prediction, we propose a new intelligent churn prediction system for telecom, named FW-ECP. The novelty of the FW-ECP lies in its ability to combine both filterand wrapper-based feature selection as well as exploit the learning capability of an ensemble classifier built using diverse base classifiers. In the filter phase, Particle Swarm Optimization-based undersampling and mRMR feature selection are employed to reduce the effect of imbalanced class distribution and large dimensionality. In Wrapper phase, we employ Genetic Algorithm that further discards irrelevant and redundant features. Random Forest, Rotation Forest, RotBoost and SVMs are then employed to exploit the new feature space. Finally, the ensemble classifier is constructed using both majority voting and stacking. We have tested and compared the performance of proposed FW-ECP system on two publicly available standard telecom datasets: Orange and Cell2Cell. FW-ECP takes into account both the imbalanced nature and large dimensionality of the training sets and yields better prediction performances compared with existing state-of-the-art approaches. The feature spaces for the Orange and Cell2Cell datasets are reduced to 24D and 18D, from 260D and 76D, respectively. The AUCs obtained by FW-ECP are 0.85 and 0.82 for Orange and Cell2Cell datasets, respectively.