FGTrans: a frequency-domain attention and dynamic graph neural network-based transformer for time series classification
Lei Gong, Shilin Zhou, Yi Hou, Huiling Chen, Zibo Yu · 2025
In recent years, transformer-based models for Time Series Classification (TSC) have achieved remarkable progress. However, time series often encounter distribution shifts and noise interference, which undermine the robustness of existing methods in highly dynamic and noisy environments. Transformer has become the latest model due to its ability to alleviate the computational complexity brought by long sequences via designing sparse attention or adopting decomposition mechanisms, but these approaches still require further improvements in handling volatile distributions and noise sensitivity. To address these challenges, this paper proposes a new model, Frequency-GNN-Transformer (FGTrans), which integrates frequency-domain attention into a Transformer architecture together with a dynamic graph neural network. By performing Fast Fourier Transform (FFT) to extract principal frequency components and conducting multi-scale convolution in the frequency domain, FGTrans enhances the modeling capacity for periodic patterns and noise. Meanwhile, a dynamic graph neural network is introduced in the time domain to capture dynamic dependencies across time steps. Experiments on multiple real-world time series datasets demonstrate that FGTrans achieves higher classification accuracy than existing methods (with an average accuracy increase of 2%). The source code of FGTrans is available on Github.