TSLANet: Rethinking Transformers for Time Series Representation Learning
Emadeldeen Eldele, Mohamed Ragab, Zhenghua Chen, Min Wu, Xiaoli Li
Abstract
Time series data inherently contains both short- and long-range dependencies, posing persistent challenges for analysis across diverse applications. Although Transformer-based approaches are effective at modeling long-term dependencies, they often suffer from high computational cost, sensitivity to noise, and overfitting on small datasets. To address these issues, we propose TSLANet, a lightweight and universal convolutional framework for time series analysis. The model incorporates an Adaptive Spectral Block that leverages Fourier analysis to represent features in both temporal and frequency domains, enabling effective modeling of local and global dependencies while suppressing noise through adaptive thresholding. Furthermore, an Interactive Convolution Block is introduced, and self-supervised learning strategies are employed to enhance temporal representation learning and improve robustness across datasets. Extensive experiments demonstrate that TSLANet achieves state-of-the-art performance in forecasting, classification, and anomaly detection tasks, exhibiting strong adaptability and stability across varying noise intensities and data scales.