A novel profit maximizing metric for measuring classification performance of customer churn prediction models
Authors: Thomas Verbraken, Frank Goethals, Wouter Verbeke, and Bart Baesens
Appears in: Decision Support Systems
Keywords: Network-based classifiers, e-commerce acceptance, relational classifiers, collective inference procedures, homophily, cohesion
The goal of this paper is to identify a new way to predict whether a specific person believes buying online is appropriate for a specific product. By analyzing data that was gathered through a survey, we show that knowledge of a person's social network can be helpful to predict that person's e-commerce acceptance for different products. Our experimental setup is interesting for companies because (1) knowledge about only a small number of connections of potential customers is needed; (2) knowing the intensity of the relation is not necessary, and (3) data concerning variables such as age, gender and whether one likes working with the PC is not needed. Hence, companies can rely on publicly available data on their customers' social ties. Network-based classifiers tend to perform especially well for highly durable goods and for services for which few customers think it is appropriate to reserve them online.