Repo Status:
Implementation:
🚩 News (2024.05.14) Compatible with MPS backend, TEFN can be trained by .
This is the official code implementation project for paper "Time Evidence Fusion Network: Multi-source View in
Long-Term Time Series Forecasting". The code implementation refers
to .
Thanks very much
for
's contribution to this project.
The Time Evidence Fusion Network (TEFN) is a groundbreaking deep learning model designed for long-term time series
forecasting. It integrates the principles of information fusion and evidence theory to achieve superior performance in
real-world applications where timely predictions are crucial. TEFN introduces the Basic Probability Assignment (BPA)
Module, leveraging fuzzy theory, and the Time Evidence Fusion Network to enhance prediction accuracy, stability, and
interpretability.
- Information Fusion Perspective: TEFN addresses time series forecasting from a unique angle, focusing on the fusion
of multi-source information to boost prediction accuracy.
- BPA Module: At its core, TEFN incorporates a BPA Module that maps diverse information sources to probability
distributions related to the target outcome. This module exploits the interpretability of evidence theory, using fuzzy
membership functions to represent uncertainty in predictions.
- Interpretability: Due to its roots in fuzzy logic, TEFN provides clear insights into the decision-making process,
enhancing model explainability.
- State-of-the-Art Performance: TEFN demonstrates competitive results, with prediction errors comparable to leading
models like PatchTST, while maintaining high efficiency and requiring fewer parameters than complex models such as
Dlinear.
- Robustness and Stability: The model showcases resilience to hyperparameter tuning, exhibiting minimal fluctuations
even under random selections, ensuring consistent performance across various settings.
- Efficiency: With optimized training times and a compact model footprint, TEFN is particularly suitable for
resource-constrained environments.
Clone the repository:
git clone https://github.com/ztxtech/Time-Evidence-Fusion-Network.git
cd Time-Evidence-Fusion-Network
pip install -r requirements.txt
You can obtain datasets
from
or
,
Then place the downloaded data in the folder
./dataset
.
- Modify the specific configuration file in
./run_config.py
.
config_path = '{your chosen config file path}'
- Run
./run_config.py
directly.
python run_config.py
- Find required configuration file
*.json
in./configs
. - Modify
*.json
file.
{
# ...
# Nvidia CUDA Device {0}
# 'gpu': 0
# Apple MPS Device
# 'gpu': 'mps'
# ...
}
Other related operations refer
to .
If you find TEFN useful in your research, please cite our work as per the citation.
@misc{TEFN,
title={Time Evidence Fusion Network: Multi-source View in Long-Term Time Series Forecasting},
author={Tianxiang Zhan and Yuanpeng He and Zhen Li and Yong Deng},
year={2024},
journal={arXiv}
}
We appreciate the following GitHub repos a lot for their valuable code and efforts.
From Time Series Library
This library is constructed based on the following repos:
- Forecasting: https://github.com/thuml/Autoformer.
All the experiment datasets are public, and we obtain them from the following links:
-
Long-term Forecasting and Imputation: https://github.com/thuml/Autoformer.
-
Short-term Forecasting: https://github.com/ServiceNow/N-BEATS.
If you have any questions or suggestions, feel free to contact:
- (Primary) Tianxiang Zhan ([email protected])
- Yuanpeng He ([email protected])
Or describe it in Issues.