Explainable Graph Neural Networks for Optimized Allocation in Spatio‐Temporal Bike Sharing Demand Prediction

ABSTRACT This research proposes a novel approach to improving the allocation strategies in spatio‐temporal bike sharing demand prediction by enhancing Explainable Graph Neural Networks (X‐GNNs). Traditional models often struggle to capture the intricate relationships between spatially distributed bike stations and their temporal demand patterns. In response to this challenge, our proposed framework employs Graph Neural Networks (GNNs) to model the complex interactions within the bike sharing network. Additionally, we optimize interpretability and decision‐making by incorporating explainability mechanisms into the model. The X‐GNN architecture integrates attention mechanisms and graph attention networks to effectively capture spatial dependencies and temporal dynamics. This not only enables accurate prediction of short‐term demand but also provides a clear understanding of the factors influencing the predictions. The attention mechanisms allow the model to focus on crucial nodes and temporal patterns, offering insights into the spatial and temporal features that contribute most significantly to the demand fluctuations. The explainability aspect of the model facilitates transparency in decision‐making processes related to resource allocation and station management. Importantly, the explainability aspect of the model facilitates transparency in decision‐making processes related to resource allocation and station management. By revealing the underlying drivers of demand—such as weather conditions, peak usage hours, or specific high‐demand locations—decision‐makers can deploy bikes more efficiently, plan for demand surges, and design proactive redistribution strategies. This bridges the gap between advanced AI models and practical, real‐world applications in urban mobility systems. To validate the effectiveness of our proposed framework, extensive experiments were conducted on spatio‐temporal bike sharing datasets. The dataset contains weather information, the number of bikes rented per hour, and date information. The results demonstrate superior prediction accuracy compared to baseline models, along with optimized interpretability.