File size: 2,892 Bytes
fe15010 1734d07 fe15010 a60a695 1734d07 a60a695 1734d07 a60a695 1734d07 a60a695 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 | ---
license: apache-2.0
pipeline_tag: other
---
# POST: Prior-Observation Adversarial Learning of Spatio-Temporal Associations for Multivariate Time Series Anomaly Detection
This repository contains the official PyTorch implementation and pre-trained checkpoints for the paper [POST: Prior-Observation Adversarial Learning of Spatio-Temporal Associations for Multivariate Time Series Anomaly Detection](https://huggingface.co/papers/2605.18128).
## 💡 Highlights
- **Novel Framework**: We propose POST, a model that unifies spatial and temporal anomaly modeling under a joint prior–observation adversarial learning framework.
- **Anomaly Localization**: The adversarial optimization not only improves time-wise detection but also enables the model to localize anomalies to specific channels.
- **SMD+ Dataset**: Construction of a dataset featuring precise channel-wise annotations, serving as a generic testbed for anomaly localization in MTS.
- **SOTA Performance**: Extensive experiments demonstrate that POST consistently outperforms state-of-the-art (SOTA) baselines in standard anomaly detection tasks.
## 🛠️ Resources
- **Code**: [GitHub Repository](https://github.com/anocodetest1/POST)
- **Paper**: [huggingface.co/papers/2605.18128](https://huggingface.co/papers/2605.18128)
## 🚀 Get Started
### Evaluation
To evaluate the trained models on the benchmarks, run the following scripts. Ensure you have the datasets prepared and the pre-trained weights placed into the `./checkpoints/` folder (named as `<dataset_name>_checkpoint.pth`).
```bash
# SMD
python main.py --anomaly_ratio 0.5 --batch_size 64 --mode test --dataset SMD --data_path dataset/SMD --input_c 38 --output_c 38
# MSL
python main.py --anomaly_ratio 1. --batch_size 64 --mode test --dataset MSL --data_path dataset/MSL --input_c 55 --output_c 55
# SMAP
python main.py --anomaly_ratio 1. --batch_size 64 --mode test --dataset SMAP --data_path dataset/SMAP --input_c 25 --output_c 25
# SWaT
python main.py --anomaly_ratio 1. --win_size 150 --batch_size 64 --mode test --dataset SWaT --data_path dataset/SWaT --input_c 51 --output_c 51
# PSM
python main.py --anomaly_ratio 1. --batch_size 64 --mode test --dataset PSM --data_path dataset/PSM --input_c 25 --output_c 25
# SMD+
python main_channel.py --anomaly_ratio 0.5 --batch_size 64 --mode test --dataset SMD+ --data_path dataset/SMD+ --input_c 38 --output_c 38
```
## Suggested citation
Please consider citing our work:
```bibtex
@misc{zhang2026postpriorobservationadversariallearning,
title={POST: Prior-Observation Adversarial Learning of Spatio-Temporal Associations for Multivariate Time Series Anomaly Detection},
author={Suofei Zhang and Yaxuan Zheng and Haifeng Hu},
year={2026},
eprint={2605.18128},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2605.18128},
}
``` |