Real-time detection of unusual customer behavior in retail using lstm autoencoders

Abstract

Personal customer care is one of the advantages of physical retail over its online competition, but cost pressure forces retailers to deploy staff as efficiently as possible resulting in a trend of staff reduction. For staff and managers it becomes harder to keep track of what is happening in a store. Situations that would benefit from intervention like cases of aimless customers, lost children or shoplifting go unnoticed. To this end, real-time tracking systems can provide managers with live data on the current in-store situation, but analysis methods are necessary to actually interpret these data. In particular, anomaly detection can highlight unusual situations that require a closer look. Unfortunately, existing algorithms are not well-suited for a retail scenario as they were designed for different use cases or are slow to compute. To resolve this, we investigate the use of long short-term memory autoencoders, which have recently shown to be successful in related scenarios, for real-time detection of unusual customer behavior. As we demonstrate, autoencoders reconcile the precision of reliable methods that have poor performance with a speed suitable for practical use.