SSDA: Spectral and Structural Dual Adaptation for Time Series Forecasting

Overview

SSDA is a vision-based time series forecasting framework that adapts Large Vision Models (LVMs) through dual adaptation mechanisms. Unlike prior work that directly applies pre-trained vision models to rendered time series, SSDA explicitly identifies and bridges two fundamental modality gaps:

• Spectral Gap: Mismatch in frequency-domain statistics between time series images and natural images
• Structural Gap: Distortion of temporal relationships caused by 1D-to-2D reshaping

To address these issues, SSDA introduces a dual-branch architecture:

• A Spectral Branch that aligns frequency statistics at the data level
• A Structural Branch that restores temporal structure at the model level

The two branches are adaptively fused to produce the final prediction.

---

Motivation

Recent work shows that rendering time series as images allows LVMs (e.g., MAE) to achieve strong forecasting performance. However, this paradigm relies on a strong assumption:

Rendered time series images are sufficiently similar to natural images.

SSDA revisits this assumption and shows:

• Time series images exhibit different power spectrum distributions (lower α)

• 2D reshaping introduces:

  • Spurious spatial adjacency
  • Broken temporal continuity

These issues fundamentally limit the transferability of pre-trained vision models.

---

Architecture

Figure 2: Overall architecture of SSDA.

---

Installation

'''bash git clone https://anonymous.4open.science/r/SSDA-8C5B cd SSDA

pip install -r requirements.txt '''

---

Quick Start

Training

'''bash python -u run.py
--task_name long_term_forecast
--is_training 1
--root_path /remote-home/data/ETT/
--data_path ETTh1.csv
--model_id ETTh1_1440_96
--model SSDA
--data ETTh1
--features M
--seq_len 1440
--label_len 48
--pred_len 96
--e_layers 2
--d_layers 1
--factor 1
--enc_in 7
--dec_in 7
--c_out 7
--learning_rate 0.000002
--des 'Exp'
--n_heads 2
--itr 1
--periodicity 24
--percent 100 '''

---

Model Parameters

Parameter	Description	Default
seq_len	Input sequence length	1440
label_len	Start token length	48
pred_len	Prediction length	96
d_model	Hidden dimension	512
n_heads	Attention heads	8
e_layers	Encoder layers	2
periodicity	Image reshape periodicity	24
r	LoRA rank	4
lora_alpha	LoRA scaling	16
residual_weight (λ)	Spectral residual weight	0.05

---

Datasets

SSDA supports the following datasets:

• ETT (ETTh1, ETTh2, ETTm1, ETTm2)
• Weather
• Electricity
• Traffic

---

Results

SSDA achieves:

• State-of-the-art performance across 7 benchmarks
• 48 first-place results in full-shot experiments
• ~7.8% MSE reduction over strongest baselines

---