SeesawNet

Official source code for the IJCAI 2026 accepted paper:

SeesawNet: Towards Non-stationary Time Series Forecasting with Balanced Modeling of Common and Specific Dependencies

Authors

Hao Li¹, Lu Zhang^2,, Liu Chong¹, Yankai Chen³, Pengyang Wang⁴, Yingjie Zhou^1,

¹Sichuan University
²Chengdu University of Information Technology
³McGill University
⁴University of Macau
^*Corresponding authors

Overview

SeesawNet is a PyTorch implementation for long-term time series forecasting under non-stationary settings. The repository contains the main SeesawNet model, ablation variants, experiment code, and shell scripts for the benchmark datasets used in long-term forecasting.

Requirements

The code was prepared for Python-based PyTorch training. Core dependencies include:

• Python 3.9+
• PyTorch
• NumPy
• pandas
• scikit-learn
• SciPy
• matplotlib
• tqdm
• einops

Install dependencies with:

pip install -r requirements.txt

Install the PyTorch build that matches your CUDA version if you plan to train on GPU.

Dataset Preparation

Download the benchmark datasets and place them under dataset/.

The datasets can be obtained from the dataset links provided by the iTransformer repository:

• iTransformer repository: https://github.com/thuml/iTransformer
• Google Drive dataset archive: https://drive.google.com/file/d/1l51QsKvQPcqILT3DwfjCgx8Dsg2rpjot/view?usp=drive_link

Expected examples:

dataset/
  ETT-small/
    ETTh1.csv
    ETTh2.csv
    ETTm1.csv
    ETTm2.csv
  electricity/
    electricity.csv
  exchange_rate/
    exchange_rate.csv
  illness/
    national_illness.csv
  Solar/
    solar_AL.txt
  weather/
    weather.csv

Usage

Clone the repository and enter the project directory:

git clone https://github.com/dreamone-Lee/SeesawNet.git
cd SeesawNet

Install dependencies:

pip install -r requirements.txt

Run the full SeesawNet benchmark script:

sh scripts/SeesawNet.sh

You can also run a single dataset script from scripts/run/, for example:

sh scripts/run/etth1.sh SeesawNet SeesawNet_TFMAE 0 3 TFMAE

Arguments passed to dataset scripts are:

model_name model_id gpu_id itr loss

Alternatively, run run.py directly with custom hyperparameters:

python -u run.py \
  --is_training 1 \
  --model_id SeesawNet_ETTh1 \
  --model SeesawNet \
  --data ETTh1 \
  --root_path ./dataset/ETT-small/ \
  --data_path ETTh1.csv \
  --seq_len 720 \
  --pred_len 96 \
  --enc_in 7 \
  --patch_len 32 \
  --stride 16 \
  --pd_layers 1 \
  --cr_layers 2 \
  --d_model 128 \
  --d_ff 256 \
  --n_heads 16 \
  --dropout 0.4 \
  --down_sample_rate 1.0 \
  --batch_size 256 \
  --learning_rate 0.0001 \
  --loss TFMAE \
  --itr 3

Repository Structure

data_provider/   Dataset loading and preprocessing
experiments/     Training, validation, and testing loops
layers/          SeesawNet backbone layers and attention modules
model/           SeesawNet and ablation model definitions
scripts/         Reproduction scripts
utils/           Metrics, losses, time features, and training utilities
run.py           Main experiment entry point

Outputs

Training creates runtime artifacts such as checkpoints, logs, CSV summaries, and test visualizations. These files are ignored by Git by default:

checkpoints/
logs/
csv_results/
test_results/
results/

Citation

If you use this repository, please cite:

@inproceedings{seesawnet2026,
  title     = {SeesawNet: Towards Non-stationary Time Series Forecasting with Balanced Modeling of Common and Specific Dependencies},
  author    = {Li, Hao and Zhang, Lu and Chong, Liu and Chen, Yankai and Wang, Pengyang and Zhou, Yingjie},
  booktitle = {Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence},
  year      = {2026}
}