DWAFM: Dynamic Weighted Graph Structure Embedding Integrated with Attention and Frequency-Domain MLPs for Traffic Forecasting

Sen Shi; Zhichao Zhang; Yangfan He

cs.LGMar 1, 2026

DWAFM: Dynamic Weighted Graph Structure Embedding Integrated with Attention and Frequency-Domain MLPs for Traffic Forecasting

Authors: Sen Shi, Zhichao Zhang, Yangfan He

Organizations: School of Mathematics and Statistics, Nanjing University of Information Science and Technology, Nanjing 210044, China · Hubei Key Laboratory of Applied Mathematics, Hubei University, Wuhan 430062, China · Key Laboratory of System Control and Information Processing, Ministry of Education, Shanghai Jiao Tong University, Shanghai 200240, China · School of Communication and Artificial Intelligence, School of Integrated Circuits, Nanjing Institute of Technology, Nanjing 211167, China ·

Abstract

Accurate traffic prediction is a key task for intelligent transportation systems. The core difficulty lies in accurately modeling the complex spatial-temporal dependencies in traffic data. In recent years, improvements in network architecture have failed to bring significant performance enhancements, while embedding technology has shown great potential. However, existing embedding methods often ignore graph structure information or rely solely on static graph structures, making it difficult to effectively capture the dynamic associations between nodes that evolve over time. To address this issue, this letter proposes a novel dynamic weighted graph structure (DWGS) embedding method, which relies on a graph structure that can truly reflect the changes in the strength of dynamic associations between nodes over time. By first combining the DWGS embedding with the spatial-temporal adaptive embedding, as well as the temporal embedding and feature embedding, and then integrating attention and frequency-domain multi-layer perceptrons (MLPs), we design a novel traffic prediction model, termed the DWGS embedding integrated with attention and frequency-domain MLPs (DWAFM). Experiments on five real-world traffic datasets show that the DWAFM achieves better prediction performance than some state-of-the-arts.

Figures & tables

Fig. 1: The overview of the proposed DWAFM.

Fig. 2: Learning mechanism of dynamic weighted adjacency matrix.

Datasets	Metric	HI	STGCN	GWNet	StemGNN	STNorm	STID	DGCRN	MegaCRN	STWave	STAEformer	DFDGCN	DWAFM
PEMS08	MAE	$34.66$	$16.86$	$14.42$	$16.08$	$15.47$	$14.21$	$14.90$	$14.87$	$13.78$	$\underline{13.69}$	$14.21$	$\mathbf{13.57}$
	RMSE	$50.54$	$26.10$	$23.38$	$25.49$	$25.13$	$23.36$	$23.76$	$23.87$	$23.68$	$\underline{23.06}$	$23.45$	$\mathbf{22.94}$
	MAPE	$21.63$	$11.62$	$9.33$	$10.19$	$9.93$	$9.38$	$10.17$	$9.58$	$9.36$	$\underline{9.03}$	$9.37$	$\mathbf{8.92}$
PEMS04	MAE	$42.35$	$19.89$	$18.76$	$21.01$	$18.90$	$18.46$	$18.96$	$18.83$	$18.28$	$\underline{18.21}$	$18.62$	$\mathbf{18.14}$
	RMSE	$61.66$	$31.59$	$30.19$	$32.96$	$30.50$	$\underline{29.98}$	$30.77$	$30.54$	$\underline{29.98}$	$30.57$	$30.01$	$\mathbf{29.72}$
	MAPE	$29.22$	$13.61$	$13.06$	$14.51$	$13.00$	$12.55$	$12.87$	$12.83$	$\underline{12.25}$	$\mathbf{12.21}$	$12.94$	$12.37$

TABLE I: Performance Comparison of Baselines and Proposed DWAFM on Five Popular used Real-world Traffic Datasets.

Fig. 3: Ablation study on PEMS08 and PEMSD7(M).

Fig. 4: Hyperparameter study and computational cost analysis on PEMS08.

Fig. 5: Visualization of learned dynamic weighted adjacency matrix.

Models	Metric (PEMS08)				Metric (PEMSD7(M))
Models	MAE	RMSE	MAPE	Time/s	MAE	RMSE	MAPE	Time/s
CNN	$13.83$	$23.70$	$9.16$	$\mathbf{26.63}$	$2.60$	$5.27$	$6.51$	$\mathbf{19.48}$
Attention	$13.85$	$23.54$	$9.14$	$26.85$	$2.73$	$5.26$	$6.91$	$23.40$
Fre-MLPs	$\mathbf{13.57}$	$\mathbf{22.94}$	$\mathbf{8.92}$	$26.71$	$\mathbf{2.59}$	$\mathbf{5.24}$	$\mathbf{6.40}$	$19.53$

TABLE II: Performance Comparison of Different Temporal Modeling Modules.

Fig. 6: Visualization of real traffic forecasting results.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Datasets	Nodes	Length	Sample Rate	Type
PEMS08	$170$	$17856$	$5$ min	Traffic flow
PEMS04	$307$	$16992$	$5$ min	Traffic flow
PEMS03	$358$	$26208$	$5$ min	Traffic flow
PEMSD7(M)	$228$	$12672$	$5$ min	Traffic speed
PEMSD7(L)	$1026$	$12672$	$5$ min	Traffic speed

Appendix

TABLE I : Statistics of Datasets.

Models	Pre-defined	Graph	S-T	Core Architecture
Models	Graph	Dynamicity	Embed.	Spatial	Temporal
STGCN	$\checkmark$	Static	-	GCN	CNN
GWNet	$\checkmark$	Static	-	GCN	CNN
StemGNN	-	Static	-	GCN	CNN
STID	-	-	$\checkmark$	MLP	MLP
DGCRN	$\checkmark$	Dynamic	-	GCN	RNN

Appendix

TABLE II : Comparison of Graph Construction and Spatial-Temporal Modules Among Different Models.

Fig. 1 : Hyperparameter study and computational cost analysis on PEMSD7(M) and PEMS04.

Methods	Metric (PEMSD7(M))			Metric (PEMS08)
Methods	MAE	RMSE	MAPE	MAE	RMSE	MAPE
Xavier_uniform	$2.59$	$5.24$	$6.40$	$13.57$	$22.94$	$8.92$
Kaiming_uniform	$2.59$	$5.25$	$6.40$	$13.59$	$22.98$	$8.94$
Orthogonal	$2.59$	$5.25$	$6.43$	$13.59$	$23.00$	$8.92$
Xavier_normal	$2.59$	$5.24$	$6.42$	$13.54$	$22.95$	$8.90$
Mean $\pm$ Std	$\mathbf{2.59\pm 0.00}$	$\mathbf{5.25\pm 0.01}$	$\mathbf{6.41\pm 0.02}$	$\mathbf{13.57\pm 0.02}$	$\mathbf{22.97\pm 0.03}$	$\mathbf{8.92\pm 0.02}$

Appendix

TABLE III : Performance Comparison under Different Node Embedding Initialization Methods.

Traffic Forecasting

Traffic forecasting predicts future flows, speeds, demand, travel times, and congestion from spatial, temporal, and contextual data.

See all 87 papers

A Dynamic Fusion Large Language Model for Traffic Flow Prediction
Sep 10, 2026Xue Qiu, Jianli XiaoIntelligent Transportation Systems Traffic Flow Estimation
School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai, China
ADMFormer: An Adaptive-Decomposition Transformer with Time-Varying Masked Spatial Attention for Traffic Forecasting
May 25, 2026Ruiwen Gu, Qitai Tan, Yahao Liu +1Attention Mechanisms Spatiotemporal Forecasting
Shenzhen Ubiquitous Data Enabling Key Lab Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · School of Computer Science and Engineering University of Electronic Science and Technology of China, Chengdu, China
A Global-Local Graph Attention Network for Traffic Forecasting
May 16, 2026Tianchi ZhangIntelligent Transportation Systems Spatiotemporal Forecasting

Sep 10, 2026J/K move · Enter open · S save

A Dynamic Fusion Large Language Model for Traffic Flow Prediction

Xue Qiu, Jianli Xiao

School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai, China

Intelligent Transportation Systems Traffic Flow Estimation Multi-Scale Feature Fusion LLM-Based Time Series Forecasting Spatiotemporal GNNs

Open paper AskPDF arXiv

Traffic flow prediction is a core supporting technology for intelligent transportation systems. It uses historical data to infer future traffic dynamics in specific areas, thereby helping to alleviate congestion and improve resource allocation efficiency. Traditional neural networks struggle to break through accuracy limits due to their reliance on singular feature modeling, while large language models (LLMs) suffer from insufficient capture of spatial topological information and mining spatiotemporal correlation. This study proposes a Dynamic Fusion Large Language Model (DF-LLM) for traffic flow prediction. The model incorporates three core components: spatiotemporal embedding module, spatiotemporal fusion module, and LLM backbone. The spatiotemporal embedding module enables synergistic representation of multi-scale spatiotemporal features. The spatiotemporal fusion module integrates spatial topology and dynamic dependencies via graph convolution. The LLM backbone adopts a differentiated parameter adaptation strategy to balance training efficiency and traffic data adaptability. Additionally, it introduces a context aggregation attention module to strengthens global dependencies. More importantly, the LLM backbone takes the residual connections to mitigate the gradient vanishing in deep networks. Experiments show that DF-LLM has achieved better performance by comparing the metrics on all the four datasets.

Abstract

Traffic Forecasting

Explore similar work

A Dynamic Fusion Large Language Model for Traffic Flow Prediction

ADMFormer: An Adaptive-Decomposition Transformer with Time-Varying Masked Spatial Attention for Traffic Forecasting

A Global-Local Graph Attention Network for Traffic Forecasting