SSDA: Dual Adaptation Bridges Spectral and Structural Gaps in Vision-Based Time Series Forecasting
A new method called SSDA (Spectral-Structural Dual Adaptation) addresses fundamental limitations in using large vision models (LVMs) for time series forecasting. Researchers identified two key gaps when rendering temporal data as images: spectral gap (rendered images have shallower power spectrum than natural images) and structural gap (reshaping 1D sequences into 2D grids creates spurious spatial adjacencies and breaks temporal continuities). SSDA uses a dual-branch network to bridge these gaps, enabling better transfer of pre-trained LVM knowledge to time series tasks. The work is published as arXiv:2605.12550.
Key facts
- SSDA stands for Spectral-Structural Dual Adaptation
- Large vision models have been used for time series forecasting by rendering data as images
- Two gaps identified: spectral gap and structural gap
- Spectral gap: rendered images have shallower power spectrum than natural images
- Structural gap: 1D to 2D reshaping creates spurious adjacencies and breaks temporal continuities
- SSDA uses a dual-branch network to address both gaps
- Published on arXiv with ID 2605.12550
- Announcement type: cross
Entities
Institutions
- arXiv