Time Slot Selection (Retail)

Time Slot Selection in e-commerce is the proof of concept of TRUST’s effectiveness in a prescriptive setting within the retail sector, bringing together INESC TEC, LTPlabs, and SONAE MC. Sonae MC is the largest Portuguese food retailer and has an e-commerce operation that, similarly to other retailers, has been boosted by the COVID pandemic. It is, thus, of the utmost importance to achieve operational efficiency without disregarding customer satisfaction.

The current operational inefficiencies in the e-commerce operation come mainly from customers’ preferences, which are clustered around the same delivery window, causing high operational loads. These operational peaks require an oversized operational capacity and result in a below-target service level. The misaligned customer preferences and supply chain objectives call for a solution compromise.

Time slot management includes defining the available slots (time windows) that customers may pick for the delivery of their order and the pricing of each slot. Time slot selection may pursue a static approach, where slots and pricing decisions are defined beforehand and do not change during the operation; or dynamically, where the decisions are adjusted based on real-time orders, taking into consideration the current operational load of each time slot and the customers’ geographic proximity. Pricing decisions can also take advantage of customers’ preferences.

The objective of this use case is to apply TRUST to dynamically select delivery time slots and corresponding pricing for last-mile delivery, preserving customer satisfaction while increasing the operational margins and efficiency. Providing explanations to the prescriptive engine’s output is key in ensuring customers are handled fairly and decision-makers fully adopt the data-driven recommendations.

In this project, a symbolic expression will score elligible time slot panels, based on variables such as customer preference, location, among others, where the highest scoring panel will be the chosen, tailored panel to be presented to each customer. To train this symbolic expression, genetic programming will be used, where expressions that generate the most profit during a simulation of real past orders are deemed the fittest.

This symbolic expression, combined with experts’ knowledge, increases the comprehension of the dynamic pricing problem. TRUST will enable two types of interactions: At the solution level, by easing cognition while interpreting a particular solution, and at the model level, by trading-off different factors such as performance, explainability, and fairness and favor one or another, depending on the circumstances.

A prescriptive, genetic programming-based methodology within the TRUST-AI framework has been successfully implemented, which dynamically sets time slot prices via a symbolic expression. This approach has enabled iterative learning loops which incorporate human expertise into the generation of the expressions. These advancements demonstrate the successful integration and practical application of the above mentioned methodology in an online retail context.