AMGST: Adaptive multi-graph convolution and spatiotemporal attention network for traffic forecasting

RESEARCH ARTICLE AMGST: Adaptive multi-graph convolution and spatiotemporal attention network for traffic forecasting Pei Shi1,2, Qixiang Lu3, Jiahui Chen1, Xiaoliu Lv1, Lu Zhang Jiadong Sun1,3 *, Liang Kuang4, 1 1 School of IoT Engineering, Wuxi University, Wuxi, China, 2 Jiangsu Engineering Research Center of Hyperconvergence application and security of IoT devices, Wuxi, China, 3 School of Electronics & Information Engineering, Nanjing University of Information Science & Technology, Nanjing, China, 4 Jiangsu Vocational College of Information Technology, Wuxi, Hebei, China * Abstract OPEN ACCESS Citation: Shi P, Lu Q, Chen J, Lv X, Zhang L, Kuang L, et al. (2026) AMGST: Adaptive multigraph convolution and spatiotemporal attention network for traffic forecasting. PLoS One 21(6): e0342235. https://doi.org/10.1371/journal. pone.0342235 Editor: Guangyin Jin, National University of Defense Technology, CHINA Received: August 27, 2025 Accepted: January 20, 2026 Published: June 4, 2026 Copyright: © 2026 Shi et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data availability statement: All data underlying the findings of this study are publicly available without restriction. The dataset used in this study is the same as that used in the published works on DCRNN and STSGCN. It can be accessed from the STSGCN GitHub repository (https://github.com/Davidham3/STSGCN) and Traffic forecasting is crucial for optimizing traffic management and control strategies. As a powerful approach for analyzing and mining graph-structured data, graph convolution has shown great potential in traffic prediction. However, it still struggles to fully capture global spatial correlations and long-term dynamic temporal dependencies inherent in spatiotemporal traffic patterns. Moreover, the quality of the graph structure directly affects the extraction of these correlations. To address these challenges, we propose AMGST, an Adaptive Multi-Graph Convolution and Spatiotemporal Multi-Head Self-Attention Network for traffic forecasting. AMGST integrates an Adaptive Spatiotemporal Embedding (ASTE) generator, a multi-graph diffusion convolution module, and a spatiotemporal attention mechanism. First, dynamic spatiotemporal representations are generated using the ASTE module. Then, the multi-graph diffusion convolution leverages both a maximum mutual information coefficient matrix and an adaptive matrix to extract fine-grained spatial features. A global spatial attention mechanism is applied to capture dynamic spatial correlations, while a temporal attention module models nonlinear temporal dependencies. Experimental results on four public traffic datasets, including both speed and flow measurements, demonstrate that AMGST consistently surpasses the baselines, confirming its effectiveness in providing accurate traffic forecasts. 1. Introduction With the rapid advancement of urbanization, traffic congestion has emerged as a pervasive issue, significantly impacting people’s daily lives and work. Traffic data is a typical spatiotemporal time series. Accurate traffic prediction can significantly enhance the optimal control and scheduling of urban transportation, thereby PLOS One | https://doi.org/10.1371/journal.pone.0342235 June 4, 2026 1 / 22 the DCRNN GitHub repository (https://github. com/liyaguang/DCRNN). Funding: The Jiangsu Province Natural Science Project of Institution (Grant No.21KJB520020), Wuxi Science and Technology Plan Project (Grant No. K20221044), the Qing Lan Project of Jiangsu Province, Wuxi Science and Technology Plan Project (K20231011), Wuxi “Xishan Talent Plan” Innovation leader talent Project (2022xsyc002), Excellent Science and Technology Innovation Team of Jiangsu Province Universities (Real-time Industrial Internet of Things). Competing interests: The authors have declared that no competing interests exist. improving the convenience of daily activities. In traffic prediction research, it is essential to consider both the temporal and spatial aspects of traffic data. Graph Convolutional Networks (GCNs) are widely used in this field due to their powerful capability to extract features from graph-structured data. Traditional GCN-based prediction methods typically rely on adjacency matrices constructed from physical distances or connectivity to extract spatial information. [1–3] However, these methods are limited in their ability to capture only local spatial dependencies, as the adjacency matrix does not account for relationships between non-adjacent nodes. To address this limitation, some researchers have employed techniques such as Dynamic Time Warping (DTW) and Pearson correlation [4] to build new graph structures capable of capturing long-range spatial dependencies. Additionally, the spatial relationships between nodes are inherently dynamic. External factors such as weather, congestion, and accidents can influence traffic flow and alter spatial traffic patterns, which static adjacency matrices fail to capture. Consequently, researchers have explored dynamic graph convolution methods to extract these dynamic spatial dependencies. Although these approaches have improved prediction accuracy to some extent, they still tend to overlook fine-grained spatial information in traffic data, limiting further enhancements in prediction performance. Temporal Convolutional Networks (TCNs) are widely used to model the temporal dependencies in traffic data. TCNs expand the receptive field by stacking multiple convolutional layers. However, deep stacking may lead to issues such as over-smoothing, increased computational overhead, and gradient explosion. Wu et al. [5] employed TCN and GCN with dilated causal convolution to extract spatiotemporal dependencies in traffic data, and used Graph WaveNet to reduce the number of layers while maintaining the same receptive field. Nevertheless, TCNs are still constrained by their receptive field and are often ineffective in capturing long-term temporal dependencies effectively. To overcome these limitations, Guo et al. [6] proposed Attention-based Spatiotemporal Graph Convolutional Networks (ASTGCN), introducing an additive attention mechanism after the spatiotemporal graph convolution blocks to extract long-term spatiotemporal dependencies. Liu et al. [7] proposed a Spatiotemporal Autoencoder (ST_AE), which projects extracted spatiotemporal features into a hidden state and applies a self-attention mechanism after the encoder to achieve strong long-term prediction performance. Zheng et al. [8] proposed the Graph Multi-Attention Network (GMAN), which directly uses spatiotemporal self-attention to capture spatiotemporal features. Yan [9] introduced GECRAN, which incorporates an attention module after the Graph Convolutional Recurrent Network to capture long-period (...truncated)