A multi-gated deep graph network with attention mechanisms for taxi demand prediction

Accurate taxi demand prediction across urban road networks is critical for optimizing taxi operations and improving urban traffic management. Traditional approaches to this problem typically rely on static temporal and spatial correlations within the road network, assuming these correlations remain constant. However, taxi demand correlations are inherently dynamic, influenced by the complex and evolving patterns of passenger requests. To address this challenge, we propose MuDGN, a Multi-Gated Deep Graph Network model, designed to predict taxi demand variations across different areas of an urban road network. The MuDGN model integrates a graph multi-attention network, a graph convolutional network (GCN) layer, and a multi-gate mechanism to achieve accurate and robust predictions. The GCN layer enhances spatial feature representation, while the multi-gate mechanism, equipped with dual gating units, further improves predictive performance. Comprehensive experiments conducted on two real-world taxi demand datasets demonstrate the superiority of MuDGN over three traditional prediction models and four state-of-the-art deep graph network models in both single-period and multi-period taxi demand prediction scenarios. These results underscore the effectiveness of MuDGN in addressing the dynamic and complex nature of taxi demand forecasting.

Applied Soft Computing

Applied Soft Computing

169

1872-9681

N/A

N/A

14