The paper deals with the development and validation of a time-based recommender system for repeat purchase decisions in food retail. For this purpose, an introduction to recommender systems was given at the beginning and classical problems for grocery retail transaction data were derived. By looking at state-of-the-art solutions for integrating a temporal component into recommendations, a transition to current models in the literature was created. Two specific models were identified that deal specifically with repeat purchase decisions. These were applied to a real data set of a stationary German food retailer and discussed with respect to the previously derived problems of recommender systems. It is shown that a Bayesian Hierarchical Model with Gamma Prior and the assumption of a Poisson process following purchase events is well suited for the sparse data situation at hand.
Abstract: In the course of growing online retailing, recommendation systems have become established that derive recommendations from customers’ purchase histories. Recommending suitable food products can represent a lucrative added value for food retailers, but at the same time challenges them to make good predictions for repeated food purchases. Repeat purchase recommendations have been little explored in the literature. These predict when a product will be purchased again by a customer. This is especially important for food recommendations, since it is not the frequency of the same item in the shopping basket that is relevant for determining repeat purchase intervals, but rather their difference over time. In this paper, in addition to critically reflecting classical recommendation systems on the underlying repeat purchase context, two models for online product recommendations are derived from the literature, validated and discussed for the food context using real transaction data of a German stationary food retailer.