MICN

Code release of paper "MICN: Multi-scale Local and Global Context Modeling for Long-term Series Forecasting" (ICLR 2023 oral)

MICN achieve SOTA on six benchmarks.

Overall Architecture

As shown in Figure 1, we decompose the time series into seasonal part and trend part by Multi-scale Hybrid Decomposition. For seasonal part, we use Seasonal Prediction block to predict. For trend part, we use simple regression to predict.

Seasonal Prediction block

The seasonal part contains several different patterns after Multi-scale Hybrid Decomposition. For each pattern, we use local-global module to extract local information and global correlations.

Local-Global module

We use downsampling convolution to extract local features and isometric convolution to capture global correlations.

Get Started

pip install -r requirements.txt
Data. All the six benchmark datasets can be obtained from Google Drive or Tsinghua Cloud.
Reproducibility. We provide the experiment scripts of all benchmarks under the folder ./scripts. You can reproduce the experiments results by:

bash ./scipts/ETTm.sh
bash ./scipts/ETTh.sh
bash ./scipts/ECL.sh
bash ./scipts/Exchange.sh
bash ./scipts/Traffic.sh
bash ./scipts/WTH.sh
bash ./scipts/ILI.sh

Experiments

Main Results

Multivariate results

arch

Univariate results

arch

Model Analysis

Local-global vs. self-attetion, Auto-correlation

arch arch

Visualization

Visualization of learned trend-cyclical part prediction and seasonal part prediction.

arch

Contact

If you have any questions, please contact wanghuiqiang@stu.scu.edu.cn. Welcome to discuss together.

Citation

If you find this repo useful, please cite our paper

@article{micn,
  title={MICN: Multi-scale Local and Global Context Modeling for Long-term Series Forecasting},
  author={Huiqiang Wang and Jian Peng and Feihu Huang and Jince Wang and Junhui Chen and Yifei Xiao},
  booktitle={International Conference on Learning Representations},
  year={2023}
}