Revenue-Centered Delivery Time Presentation on Platforms:  A Spatial Neural ODE Approach

Abstract

How can e-commerce platforms present predictive information to optimize revenue while managing uncertainty in dynamic predictions? Delivery time prediction is a prime example, where inaccuracies—whether early or late—impose asymmetric costs that affect customer satisfaction and platform profitability. This study formulates a new problem for optimal delivery time presentation, explicitly accounting for the penalties of early and late deliveries through a revenue-centered loss function. To tackle the unique challenges of sporadic and episodic orders on spatial networks in this problem, we propose the Strategic Spatial Neural Ordinary Differential Equation (S^2-ODE) model, which incorporates continuous-time dynamics on spatial networks. Using real-world data from RiRiShun Logistics (RRS), a leading home appliance delivery provider in China, we demonstrate that S^2-ODE reduces revenue losses by 38.9% to 75.4% compared to state-of-the-art benchmarks in dynamic prediction, yielding savings of $0.41 to $0.53 per order. Moreover, our revenue-centered loss function achieves an additional reduction in revenue losses by  36.6% to 82.2%, translating to savings of $0.43 to $1.00 per order. We analyze revenue impact of dynamic spatial network information and continuous-time modeling, highlighting their roles in mitigating revenue losses. 

This work contributes to the platform, design science, and transportation research by introducing a revenue-centered problem formulation that  aligns predictive models with business objectives, moving beyond traditional error minimization. Moreover, our model advances deep learning and spatio-temporal prediction through network-based continuous-time modeling, enabling accurate and dynamic predictions for the irregular and sporadic observations typical of e-commerce platforms. 

Keywords: information presentation, deep learning, spatio-temporal learning, ordinary differential equations, delivery time prediction