Towards scalable taxi demand prediction

Accurate taxi demand prediction is essential for optimizing urban mobility systems across varying spatial-temporal resolutions and data conditions. Scalable taxi demand prediction refers to the capability of forecasting models to adapt to different granularities of spatial and temporal data while maintaining prediction accuracy, a critical requirement for practical urban applications ranging from fleet management to transportation planning. However, two fundamental challenges impede this scalability: data sparsity and multi-resolution forecasting requirements. Data sparsity, particularly pronounced in high-resolution predictions where numerous regions exhibit minimal activity, significantly compromises model performance. Concurrently, different urban applications necessitate predictions at varying temporal and spatial granularities, requiring models with adaptive forecasting capabilities. Our research systematically addresses these challenges by first developing solutions for data sparsity, which then enable effective multi-resolution prediction capabilities. We introduce STZIP-GNN, a Spatiotemporal Zero- Inflated Poisson Graph Neural Network that explicitly models structural zeros through a ZIP probability layer while integrating auxiliary information from crowdsensed geolocation and socioeconomic data. This approach effectively mitigates the sparse demand patterns frequently encountered in high-resolution data. Building upon this foundation, we develop SSTZIP-GNN, which incorporates an adaptive learning mechanism that dynamically adjusts forecasting strategies based on input data granularity, enabling scalable, multi-resolution prediction within a unified framework. Extensive experiments conducted on 130 million real-world mobility records demonstrate that our proposed models significantly outperform existing baselines. SSTZIP- GNN achieves up to 53.3% RMSE reduction and 46.3% lower computational costs, exhibiting superior robustness and scalability across diverse urban conditions. These advancements provide municipalities with a unified tool for demand-responsive fleet management, dynamic pricing, and sustainable mobility planning across heterogeneous urban landscapes. Our research contributes to the development of more efficient and adaptable intelligent transportation systems.