README.md

CMA: A Unified Contextual Meta-Adaptation Methodology for Time-Series Denoising and Prediction [Paper]

Accepted in KDD 2025

Haiqi Jiang, Ying Ding, Chenjie Pan, Aimin Huang, Rui Chen, Chenyou Fan

Requirements and Installation

conda create -n cma python==3.10
conda activate cma
conda install pytorch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 pytorch-cuda=12.1 -c pytorch -c nvidia
pip install einops ema-pytorch matplotlib scikit-learn scipy seaborn tqdm dm-control dm-env dm-tree mujoco gluonts

Datasets

Please download our self-collected industrial datasets (US Stock, China Stock, China Carbon Emission Data) from Drive

Please download Academic datasets (Traffic, ETTh, etc) from Autoformer official website.

Unzip them and place under ./Data/datasets, or change Config/ yaml for your customized locations.

Training CMA

Step 1: Run the code of "Timexer" or "iTransformer" to get a pretrained TSP.

Here we demonstrate iTransformer as a TSP backbone, and show example on training with Traffic data

bash scripts_itrans/iTrans_traffic.sh  

After getting model checkpoints, modify the model path: ./Check_itrans/checkpoint_traffic.pth

Step 2: Pretrain the initial CMA

python main.py --name traffic --config_file Config/traffic_dts.yaml --gpu=4 --train --milestone 1

Step 3: Train the CMA with adaptation

python main.py --name traffic --config_file Config/traffic.yaml --gpu=4 --train --milestone 10 --pretrained

Testing CMA

python main.py --name traffic --config_file Config/traffic.yaml --gpu 3 --sample 1  --milestone 20  --mode predict --pred_len 96

Citation

If you find this repo helpful in your research, please kindly cite us.

@inproceedings{jiang2025cma,
author = {Jiang, Haiqi and Ding, Ying and Pan, Chenjie and Huang, Aimin and Chen, Rui and Fan, Chenyou},
title = {CMA: A Unified Contextual Meta-Adaptation Methodology for Time-Series Denoising and Prediction},
year = {2025},
booktitle = {Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)},
}