Your next slot is... Explainable customer choice modeling for attended home delivery

In today’s competitive e-commerce landscape, retailers continuously seek innovative solutions to enhance customer experience while managing operational efficiency. One critical challenge is attended home deliveries, where customers select a delivery time slot for a specific fee. This selection process impacts both customer satisfaction and the retailer’s transportation costs, necessitating an optimized approach to pricing and scheduling.

A European online retailer faces this challenge daily, offering multiple delivery time slots with varying pricing. Understanding customer preferences in this selection process is crucial to optimizing pricing strategies and improving efficiency. The key challenges include:

• Customer Sensitivity to Pricing: Customers are highly price-conscious when selecting delivery slots.

• Operational Costs: Knowing customer preferences can be crucial to help steer their choices to more efficient slots and thus, reduce transportation expenses

• Predictive Complexity: Accurately anticipating customer choices while maintaining model interpretability for business use.

To address this challenge, we tested twomethodologies:

1. Traditional Machine Learning Models: Performant yet "black-box" models, that require explainability techniques to better understand their inner workings.
2. Genetic Programming Symbolic Expressions: This approach derives transparent mathematical expressions that model customer choices.

We treated the problem as a classification task,predicting the probability of a time slot being chosen based on customer andslot attributes. Key factors considered included:

• Pricing Features: The cost of a time slot and its comparison to other options.
• Time Slot Features: Delivery day, time of day, and proximity to the order time.
• Customer Features: Past purchasing behavior, basket value, and previous delivery preferences.
• Relationship Features: The alignment of the selected slot with historical choices.

Our comparative analysis revealed a trade-off betweenperformance and interpretability:

• Predictive Accuracy: GBM achieved the highest accuracy, correctly predicting customer selection in 29% of cases, while symbolic expressions followed closely at 24%, both significantly outperforming a naive baseline assumption (17%). This is a huge achievement, since, in average, each client sees 40 options.
• Computational Efficiency: Symbolic expressions generated predictions 17 times faster than traditional machine learning models, a crucial advantage in real-time e-commerce interactions.
• Transparency and Interpretability: While black-box models required additional analysis to extract insights, symbolic expressions directly illustrated how pricing, time slot attributes, and customer history influenced decisions.
• Business Insights: Both methods confirmed that customers are price-sensitive and prioritize earlier delivery slots. Symbolic expressions provided explicit mathematical relationships that stakeholders could use to balance price adjustments with customer preferences.

This study highlights several critical implicationsfor e-commerce retailers:

• Trust and Adoption: Transparent models increase stakeholder confidence in data-driven decision-making.
• Actionable Insights: Directly observable relationships help refine pricing strategies and operational efficiency.
• Balanced Approach: Businesses should consider using black-box models for high-stakes accuracy needs while leveraging white-box models for strategic decision-making.
• Speed Advantage: In time-sensitive applications, symbolic expressions offer faster computations, enhancing user experience.

By combining predictive accuracy withinterpretability, retailers can optimize delivery pricing strategies, improveoperational efficiency, and enhance customer satisfaction. This case studyunderscores the potential of symbolic expressions in delivering actionablebusiness insights while maintaining a high level of model transparency. Ase-commerce evolves, leveraging both traditional and interpretable machinelearning approaches will be crucial for sustainable growth and competitiveadvantage.