Introducing the Time Series Library (TSLib)
Overview
The Time Series Library (TSLib) is an open-source platform specifically designed for researchers in the field of deep learning, with a focus on deep time series analysis. It provides an extensive and tidy code base that facilitates the evaluation and development of advanced time series models. TSLib covers five major tasks: long- and short-term forecasting, data imputation, anomaly detection, and classification.
Key Features
- Diverse Tasks: TSLib supports a broad range of tasks necessary for effective time series analysis, from forecasting to anomaly detection.
- Comprehensive Code Base: Researchers can use existing models or contribute their own, enhancing the toolkit available for time series projects.
- State-of-the-Art Models: Inclusion of cutting-edge models such as iTransformer and TimeXer, which introduce new paradigms in time series forecasting.
Recent Updates
- Introduction of TimeXer (October 2024): This model promotes a novel "Forecasting with Exogenous Variables" paradigm, balancing practicality with computational efficiency.
- Open-Sourcing of OpenLTM (October 2024): OpenLTM marks a new approach differing from TSLib by using a pretrain-finetuning paradigm, useful for those interested in Large Time Series Models.
- Survey of Deep Time Series Models (July 2024): A comprehensive survey backed by rigorous benchmarks to guide future research endeavors.
- Inclusion of Mamba Model (April 2024): Thanks to contributions from the community, Mamba, a sequential model, has been integrated into TSLib.
Leaderboard
TSLib maintains a leaderboard showcasing top models excelling across various tasks. As of March 2024, the leading models include:
- Long-term Forecasting - Look-Back-96: TimeXer
- Short-term Forecasting: TimesNet
- Imputation and Classification: TimesNet leads across these tasks.
- Anomaly Detection: FEDformer ranks among the top models.
Usage Instructions
- Install Python 3.8: Ensure the necessary environment by executing
pip install -r requirements.txt. - Prepare Data: Obtain pre-processed datasets via provided cloud storage links and place them appropriately in the project folder.
- Model Evaluation: Use pre-defined scripts in the
./scripts/folder to test various models across different tasks like forecasting and classification. - Develop Custom Models: Easily integrate new models by following structured steps and add them to the existing framework.
Citation and Contribution
Researchers who find value in TSLib are encouraged to cite relevant papers associated with the project. The community is also invited to contribute by proposing new models or improvements, fostering collaborative progress.
Contact and Support
The project is maintained by a team of dedicated researchers and students from Tsinghua University. Users can reach out via emails provided in the documentation for any inquiries or support related to the library.
Acknowledgements
TSLib is supported by the National Key R&D Program of China and is built upon prior repositories related to forecasting, anomaly detection, and classification. Contributions from these foundational projects have been pivotal in the development of TSLib.
By providing such a robust platform, TSLib aims to push the boundaries of time series research, support innovative model development, and encourage collaborative efforts in this evolving field.