UniTime

This is the official repository of our WWW 2024 paper UniTime: A Language-Empowered Unified Model for Cross-Domain Time Series Forecasting. UniTime aims to explore the potential of utilizing a unified model for generalization across time series application domains. This code base offers an implementation to facilitate cross-domain time series learning research.

<img src='img/intro.jpg' width='750px'>

Getting Started

Requirements

Our experimental environments include Python 3.9, Pytorch 1.13.1 with CUDA 11.6, and transformers 4.31.0. To install all dependencies, please use the below command.

pip install -r requirements.txt

Datasets

The pre-processed datasets can be obtained from the link here. Then you may choose to download all_datasets.zip, place this zip file into the dataset folder, and finally unzip the file.

Running

In general, we use a csv file to indicate the executing tasks (including training and evaluations) during the cross-domain learning process. There are five columns in the file.

(1) Data: the name of a dataset, corresponding to a config file in the folder data_configs.

(2) Prediction: the prediction length.

(3) Train: the indicator for training.

(4) Valid: the indicator for validation.

(5) Test: the indicator for testing.

For example, the below command is used to train one model for the tasks listed in the file execute_list/train_all.csv. Note that the argument max_token_num should be set to a value larger than the combined number of tokens in both language instructions and time series patches.

python run.py --gpu 0 --training_list execute_list/train_all.csv --max_token_num 17

In the case of evaluating the pretrained model, please setting the argument is_training to 0 and specifying the inference tasks via inference_list.

python run.py --gpu 0 --training_list execute_list/train_all.csv --max_token_num 17 --is_training 0 --inference_list execute_list/inference_all.csv

Citation

If you find our work useful in your research, please cite:

@inproceedings{liu2024unitime,
  title={UniTime: A Language-Empowered Unified Model for Cross-Domain Time Series Forecasting},
  author={Liu, Xu and Hu, Junfeng and Li, Yuan and Diao, Shizhe and Liang, Yuxuan and Hooi, Bryan and Zimmermann, Roger},
  booktitle={Proceedings of the ACM Web Conference 2024},
  year={2024}
}

Acknowledgement

We appreciate the following github repository for sharing the valuable code base and datasets:

https://github.com/thuml/Time-Series-Library