Enhancing Customer Baskets through Assortment Optimization

This paper introduces models that support a novel system designed to optimize store selection when customers shop for baskets of items (e.g., bread, milk, snacks). Such behavior is typical in traditional grocery shopping and applicable to various e-commerce platforms. These platforms are often limited in the assortment they can provide, constrained to a fraction of the potential millions of products available (e.g., constrained to tens of thousands of products), making the problem of selection optimization both pertinent and computationally complex. To address this challenge, methods are proposed that leverage customer behavior data to identify groups of interchangeable items, enabling the formulation of customer choice models that reflect category complementarity and product substitutability. The model was initially implemented for selection optimization at one of Amazon’s fulfillment centers in early 2020. Following its success, plans were made to expand its application to multiple additional sites by the fourth quarter of the same year. Retrospective analysis shows that compared to Amazon's existing selection strategy, sales have significantly increased, with sales in specific service areas growing by 4.8%, which translates to a significant increase in annual revenue. Additionally, the deployment of the model demonstrated a 13.7% reduction in basket abandonment rates and a 13.8% increase in units per order (UPO). Efforts are ongoing to extend the model to a wider selection process, with experimental implementation within Amazon anticipated by the end of 2020.