Improving churn prediction in telecommunications using complementary fusion of multilayer features based on factorization and construction

High dimensional and unbalanced datasets are the main problems which prevent from achieving ideally churn prediction performance. Features selection is necessary to be adopted to enhance the model performance. A new predicting framework is proposed in this paper which uses complementary fusion of the multilayer features. Several subsets and new features were acquired according to feature factorization and feature construction respectively. The effective features were selected by multilayer complementary fusion which according to the contribution of feature subsets and new features. In this way, the imbalance defects of class distributions can be fixed, prediction accuracy can be improved and system stability can be reinforced. Five data mining models were applied in customer churn. Experimental results demonstrated that the method we proposed could preferably overcome the inelasticity of traditional feature selection algorithms, and more effective than those existing methods in telecommunications industry. Furthermore, we found optimal fusion with prediction model for customer churn prediction in telecommunications industry through exploring the advantages and limitations of each feature subset and prediction techniques.