Heterogeneous multi-view graph gated neural networks for real-time origin-destination matrix prediction in metro systems

Short-term origin-destination (OD) matrix prediction in metro systems faces challenges of high dimensionality, data sparsity, incomplete information, and semantic complexity. This paper proposes an effective framework called Multi-Graph Gated Neural Networks with Linear Modulation (MGGNLM) to address these challenges. We introduce distillation units to mitigate matrix dimensionality and sparsity issues, while incorporating real-time passenger flow data to handle incomplete information. The metro network is transformed into a heterogeneous graph comprising three components: a connectivity graph based on geometric location, a function-aware graph derived from GPT-2, and a mobility-pattern-aware graph constructed using Jensen-Shannon divergence. Through numerical experiments on Hangzhou and Nanjing datasets, our model demonstrates superior performance in multi-step OD demand prediction, improving WMAPE by 3.06% and 3.31% respectively compared to state-of-the-art methods. Additionally, MGGNLM exhibits exceptional performance in few-shot learning scenarios, making it particularly valuable for practical applications in metro systems.