Benchmarking Stacking Against Other Heterogeneous Ensembles in Telecom Churn Prediction

Customer churn prediction using data mining is an increasingly important concern in the highly saturated telecommunication sector. Albeit its popularity in research, few studies investigated the use of heterogeneous ensembles for this purpose. Therefore, this study evaluated and compared the performance of five grid-searched optimized base classifiers (logistic regression, decision trees, K-nearest neighbors, multilayer perceptron, and support vector machines) and their heterogeneous ensembles (stacking, grading, majority voting, weighted majority voting, and soft voting) on four different telecom datasets. The results indicate that there are significant improvements when using heterogeneous ensembles compared to single classifiers, with stacking being the most performant ensemble. The best meta-classifier for stacking was found to he a multilayer perceptron. Additionally, we identified that using probabilities as an input to the ensemble's meta-classifier, such as in soft voting and stacking variants, can increase their performance.