Code for Paper:

When Attention Fails: How Do We Mitigate Channel Noise in Multivariate Time Series?

Get Started

1. Install requirements. pip install -r requirements.txt

2. Download data. You can download the all datasets from iTransformer: datasets. All the datasets are well pre-processed and can be used easily.

3. Train the model. We provide the experiment scripts of all benchmarks under the folder ./MTSF.

4. You can use bash commands to individually run scripts in the 'scripts' folder from the command line to obtain results for individual datasets, for example, you can use the below command line to obtain the multivariate forecasting result of input-96-predict-96 on ETTh1:

   bash ./MTSF/96or12/ETTh1.sh

You can find:

The training history and results under './logs' folder.

The test results (MSE and MAE) under './test_dicts' folder.

The visualization of forecasting results under './visual/forecasting' folder.

The visualization of learning representation (Such as Global Token) under './visual/feature' folder.

The visualization of the weight under './visual/weight' folder. PS: The method used for getting the weight follows the approach outlined in An Analysis of Linear Time Series Forecasting Models.

Several examples for better understanding the proposed method in the paper.

Within the storage limits of a GitHub repository, we provide training log files of tasks input-96-predict-720 on ECL&Traffic, to assist readers in reproducing the results from the paper.