Starg2n: a spatial-temporal relevance analysis graph neural network for traffic prediction

• Published in: Pattern analysis and applications : PAA Vol. 28; no. 3
• Main Authors: Zhao, Jinbo, Xu, Xiaolong
• Format: Journal Article
• Language: English
• Published: London Springer London 01.09.2025; Springer Nature B.V
• Subjects: Artificial neural networks; Chebyshev approximation; Computer Science; Feature extraction; Graph neural networks; Intelligent transportation systems; Neural networks; Original Article; Pattern Recognition; Performance measurement; Roads & highways; Root-mean-square errors; Spatial dependencies; Spatiotemporal data; Topology; Traffic information; Transportation networks; Deep learning; Traffic prediction; Spatial-temporal data mining; Intelligent traffic system; Transportation service
• ISSN: 1433-7541; 1433-755X
• DOI: 10.1007/s10044-025-01519-5

Traffic prediction is a critical technology in intelligent transportation systems, which can provide guidance services for traffic participants and decision-making services for managers. However, traffic data is generated by the movement of road users on urban roads, which makes it constrained by road networks’ topology structures and change over time. Therefore, traffic data usually has complex spatial and temporal features, and current prediction methods have limitations in extracting these features. To this, we propose a spatial-temporal relevance analysis graph neural network, StarG2N. StarG2N constructs local spatial-temporal graphs, which are used to extract spatial-temporal features. In this process, StarG2N applies a sliding window mechanism to ensure the integrity and flexibility of feature extraction. Meanwhile, StarG2N constructs a deep neural network using K-order Chebyshev graph convolution, Gate Recurrent Unit and Multi-head Self-attention mechanism. This framework can make full use of the information of neighboring road nodes, and establish a good combination of local feature extraction and global feature analysis. StarG2N can extract spatial-temporal dependency without using different networks to model temporal and spatial dependencies separately. Extensive experiments are conducted on four real-word datasets which are all obtained from the Caltrans Performance Measurement System (PeMS). Experimental results show that StarG2N can greatly improve prediction performance compared with baselines such as T-GCN, STSGCN and OGCRNN. In terms of Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE), the maximum performance improvement of StarG2N is 5.54%, 9.14%, and 7.03% respectively.