A Stacking Approach to Direct Marketing Response Modeling

Abstract

In this work, we investigate the viability of the stacked generalization approach in predictive modeling of a direct marketing problem. We compare the performance of individual models created using different classification algorithms, and stacked ensembles of these models. The base algorithms we investigate and use to create stacked models are Neural Networks, Logistic Regression, Support Vector Machines (SVM), Naïve Bayes and Decision Tree (CART). These algorithms were selected for their popularity and good performance on similar tasks in previous studies. Using a benchmark experiment and statistical tests, we compared five single algorithm classifiers and 26 stacked ensembles of combinations these algorithms on two popular metrics: Area Under ROC Curve (AUC) and lift.  We will demonstrate a significant improvement in the AUC and lift values when the stacked generalization approach is used viz a viz the single-algorithm approach. We conclude that despite its relative obscurity in marketing applications, stacking holds great promise as an ensembling technique for direct marketing problems.