Origin–destination prediction via knowledge-enhanced hybrid learning

Abstract

This paper proposes a novel origin–destination (OD) prediction (ODP) model, namely, knowledge-enhanced hybrid spatial–temporal graph neural networks (KE-H-GNN). KE-H-GNN integrates a deep learning predictive model with traffic engineering domain knowledge and a multi-linear regression (MLR) module for incorporating external factors. Leveraging insights from the gravity model, we propose two meaningful region partitioning strategies for reducing data dimension: election districts and K-means clustering. The aggregated OD matrices and graph inputs are processed using an long short-term memory network to capture temporal correlations and a multi-graph input graph convolutional network module to capture spatial correlations. The model also employs a global–local attention module, inspired by traffic flow theory, to capture nonlinear spatial features. Finally, an MLR module was designed to quantify the relationship between OD matrices and external factors. Experiments on real-world datasets from New York and Tokyo demonstrate that KE-H-GNN outperforms all the baseline models while maintaining interpretability. Additionally, the MLR module outperformed the concatenation method for integrating external factors, regarding both performance and transparency. Moreover, the election district-based partitioning approach proved more effective and simpler for practical applications. The proposed KE-H-GNN offers an effective and interpretable solution for ODP that can be practically applied in real-world scenarios.