Calibration-CDS

This is the code for the KDD'24 paper - "Calibration of Time-Series Forecasting: Detecting and Adapting Context-Driven Distribution Shift".

How to use?

We use PatchTST and ETTh1 datasets as an example.

1. sh scripts/PatchTST/train/ETTh1.sh This is to first train the forecasting models. Here all scripts in scripts/PatchTST/train are simply copied from the original PatchTST repository, but adding an extra --run_train --run_test.
2. sh scripts/PatchTST/detection/ETTh1.sh This is to obtain the prediction residuals for calculating the Reconditionor indicators. Here --get_data_error --batch_size 1 is used.
3. python reconditionor/calc_distribution.py This is to calculate the Reconditionor indicators.
4. sh scripts/PatchTST/adaptation/ETTh1.sh This is to use SOLID to make sample-level adaptations on the forecasting models, thus making better performance. --test_train_num 1000 --run_select_with_distance --selected_data_num 10 --adapted_lr_times 10 is used to make adaptation.

Citation

@inproceedings{2024_calibration,
  title={Calibration of Time-Series Forecasting: Detecting and Adapting Context-Driven Distribution Shift},
  author={Mouxiang Chen and Lefei Shen and Han Fu and Zhuo Li and Jianling Sun and Chenghao Liu}
  booktitle={Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year={2024}
}

Star History