Towards A General Time Series Anomaly Detector with Adaptive Bottlenecks And Dual Adversarial Decoders

This code is the official PyTorch implementation of our ICLR'25 paper: Towards a General Time Series Anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders. In this paper, we produce a general time series anomaly detector DADA. By pre-training on multi-domain time series data, we achieve the goal of “one-model-for-many”, meaning a single model can perform anomaly detection on various target scenarios efficiently without domain-specific training.

If you find this project helpful, please don't forget to give it a ⭐ Star to show your support. Thank you!

🚩 News (2025.01) DADA has been accepted by ICLR 2025.

Note

The pre-training component of DADA is protected by intellectual property rights and is intended for commercial applications. As such, we are unable to disclose or release the code associated with the pre-training process.

Requirements

We run our project under Python 3.8. You can install the dependencies with the following command:

pip install -r requirements.txt

Data Preparation

Datasets can be downloaded at this link: https://drive.google.com/file/d/1QumS8bSRsLZT7u5TWLaWctDWvGnSyeRB/view?usp=drive_link.

Evaluation

As a general time series anomaly detector, DADA can perform anomaly detection on various target scenarios with out domain-specific training. You can directly test on target datasets as the following scripts:

sh ./scripts/DADA.sh

Citation

If you find this repo useful, please cite our paper:

@inproceedings{shentu2025towards,
 title     = {Towards a General Time Series Anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders},
 author    = {Shentu, Qichao and Li, Beibu and Zhao, Kai and Shu, Yang and Rao, Zhongwen and Pan, Lujia and Yang, Bin and Guo, Chenjuan},
 booktitle = {ICLR},
 year      = {2025}
}# DADA