Efficient Online Origin-Destination Prediction via a Knowledge-Driven Deep Learning Framework

Abstract

Efficient and accurate Origin-destination (OD) Prediction is critical for enhancing the efficiency of demand-responsive urban mobility systems. Existing Deep Learning (DL)-based OD Prediction models face significant challenges including extensive training time, limited adaptability to new data, and lack of explainability, which limits their real-world applications. To address these issues, this study proposes a knowledge-driven hierarchical framework, integrating a Gravity Model-Inspired Network (GIN) with a DL model (GIN-DL). The framework predicts real-time OD matrices via a hierarchical process: First, outflow prediction via a DL model. Second, flow probability estimation via GIN, which aligns with the principle of the singly constrained gravity model. The OD matrix is obtained by multiplying the predicted outflows and estimated flow probabilities. The study also introduces a two-stage training strategy combining offline pre-training using Keyframes (KF) and online fine-tuning through incremental learning. Experiments conducted on two standard benchmark datasets demonstrate that the proposed GIN-DL framework achieves substantial improvements in training time (up to 90%) while maintaining competitive prediction accuracy. Meanwhile, the integration of domain knowledge enhances explainability. The findings underscore the potential of applying the proposed framework and training strategy in real-world applications. The code is available at: https://github.com/xingzr1005/GIN-DL