Predicting Purchase Decisions Using Autoencoders in a High-Dimensional Setting

Abstract To be able to serve customers with personalized product recommendations or shopping lists, it is important for retailers to accurately predict the purchase behaviour of their customers. In this study, we propose an autoencoder-based model for next basket prediction. The goal of next basket prediction is to accurately predict the set of items a customer will buy at his next moment of purchase. Our model consists of three main steps. First, we train an autoencoder, which encodes high-dimensional baskets into low-dimensional codes. Next, this code is used to predict the code of the next basket, which is decoded in the third step to obtain basket predictions. Our model clearly outperforms the benchmark model in a wide range of performance measure and is able to deal with high dimensionality of the retailer’s assortment.