Predicting interpurchase time in a retail environment using customer-product networks: An empirical study and evaluation

Abstract

In predictive analytics and statistics, entities are frequently treated as individual actors. However, in reality this assumption is not valid. In the context of retail, similar customers will behave and thus also purchase similarly to each other. By combining their behavior in an intelligent way, based on transaction history, we can leverage these connections and improve our ability to predict purchase outcomes. As such, we can create customer-product networks from which we can deduce information on customers expressing similar purchasing behavior. This allows us to exploit their preferences and predict which products are going to be sold significantly less often. We want to use this information mainly for gaining novel marketing insights on products. For example, if customers refrain from buying products this might be due to contextual reasons such as new complements or supplements, or new nearby shops. By using these networks on data from an offline European retail corporation, we are able to boost performance of the predictive models by 6% and the identification of these specific products by 20%. This indicates that the development of customer-product graphs in retail can lead to improved marketing intelligence. To our knowledge, this is one of the first studies to use customer-product networks for predictive modeling in an offline retail setting. Furthermore, we suggest an extensive set of product and network features which can guide future researchers and practitioners in their model development.