Ensemble learning with dynamic weighting for response modeling in direct marketing

Abstract

Response modeling, a key to successful direct marketing, has become increasingly prevalent in recent years. However, it practically suffers from the difficulty of class imbalance, i.e., the number of responding (target) customers is often much smaller than that of the non-responding customers. This issue would result in a response model that is biased to the majority class, leading to the low prediction accuracy on the responding customers. In this study, we develop an Ensemble Learning with Dynamic Weighting (ELDW) approach to address the above problem. The proposed ELDW includes two stages. In the first stage, all the minority class instances are combined with different majority class instances to form a number of training subsets, and a base classifiers is trained in each subset. In the second stage, the results of the base classifiers are dynamically integrated considering two factors. The first factor is the cross entropy of neighbors in each subset, and the second factor is the feature similarity to the minority class instances. In order to evaluate the performance of ELDW, we conduct experimental studies on 10 imbalanced benchmark datasets. The results show that compared with other state-of-the-art imbalance classification algorithms, ELDW achieves higher accuracy on the minority class. Last, we apply the ELDW to a direct marketing activity of an insurance company to identify the target customers under a limited budget.