DFGNet: A dual-pathway graph neural network via frequency decomposition for spatiotemporal forecasting

Document Type

Article

Publication Date

2-1-2026

Abstract

Accurate traffic flow prediction is essential for intelligent transportation systems. Empirical analysis reveals that traffic signals inherently exhibit a dual-frequency nature: high-frequency non-stationary perturbations caused by sudden events, and low-frequency stationary trends driven by regular commuting behaviors. Most existing models fail to explicitly separate these dynamics, resulting in poor generalization under volatile traffic conditions. In this paper, we propose DFGNet, A Dual-Pathway Graph Neural Network that integrates frequency-aware decomposition with adaptive spatiotemporal learning. We design a learnable frequency separation framework to decouple raw traffic signals into non-stationary and stationary components. These are then modeled using a dual-branch architecture: a dynamic graph convolutional network (DGCN) to capture short-term structural perturbations, and a periodic-aware gated recurrent unit (PAGRU) to learn long-term temporal patterns. A lightweight adaptive fusion module dynamically balances the two branches to improve robustness and interpretability. We evaluate DFGNet on three real-world datasets: METR-LA, PEMS-BAY, and NE-BJ. Results show that our model consistently outperforms all strong baselines, achieving up to 13.95 % relative reduction in MAE and 15.38 % relative reduction in RMSE. Extensive ablation and visualization analyses further demonstrate the effectiveness of the proposed frequency-guided spatiotemporal framework. Code is available at: https://github.com/cxt1314520/DFGNet.

Keywords

Frequency decomposition, Graph neural networks, Non-stationary signal modeling, Periodic pattern modeling, Spatiotemporal forecasting, Traffic flow prediction

Publication Title

Expert Systems with Applications

ISSN

0957-4174

DOI

10.1016/j.eswa.2025.129518

Funders

Universiti Malaya, Scientific Research and Innovation Project for Graduate Students at Yunnan University, Yunnan University (KC-24249002), National Science and Technology Council (NSTC-114-2224-E-027-001, PV009-2023), National Natural Science Foundation of China (61862062), KW IPPP (RMF1506-2021)

Volume

297

Publisher

Elsevier Ltd

Share

COinS