Weakly Augmented Variational Autoencoder in Time Series Anomaly Detection

Weakly Augmented Variational Autoencoder in Time Series Anomaly Detection

Zhangkai Wu, Longbing Cao, Qi Zhang, Junxian Zhou, Hui Chen

Paper: https://arxiv.org/abs/2401.03341

Datasets

• GD
• HSS
• ECG
• TD

Requirements

• Python 3.7

• PyTorch 1.1

Quick Start

1. infoNCE Loss

# GD
# contrast
python main.py --train --seed 3 --model_name VQRAEcontrast --log_name NCE --loss_function mse  --lmbda 0.0001 --use_clip_norm --batch_size 64 --preprocessing --dataset GD

# HSS
# contrast
python main.py --train --seed 3 --model_name VQRAEcontrast --log_name NCE --loss_function mse --lmbda 0.0001 --use_clip_norm --batch_size 64 --preprocessing --dataset HSS


# TD
## contrast
python main.py --train --seed 3 --model_name VQRAEcontrast --log_name NCE --loss_function mse --lmbda 0.0001 --use_clip_norm --batch_size 64 --preprocessing --dataset TD

2. AdversirialLoss

# GD
python main.py --train --seed 3 --model_name VQRAEcontrast --log_name Discriminator --loss_function mse  --lmbda 0.0001 --use_clip_norm --batch_size 64 --preprocessing --dataset GD --discriminator

# HSS
# discriminator
python main.py --train --seed 3 --model_name VQRAEcontrast --log_name Discriminator --loss_function mse  --lmbda 0.0001 --use_clip_norm --batch_size 64 --preprocessing --dataset HSS --discriminator


# ECG
# ECG contrast
python main.py --train --seed 3 --model_name VQRAEcontrast --log_name NCE --loss_function mse --lmbda 0.0001 --use_clip_norm --batch_size 64 --preprocessing --dataset ECG
# ECG discriminator

# TD
# discriminator
python main.py --train --seed 3 --model_name VQRAEcontrast --log_name Discriminator --loss_function mse  --lmbda 0.0001 --use_clip_norm --batch_size 64 --preprocessing --dataset TD --discriminator

Citations

@article{wu2024weakly,
  title={Weakly Augmented Variational Autoencoder in Time Series Anomaly Detection},
  author={Wu, Zhangkai and Cao, Longbing and Zhang, Qi and Zhou, Junxian and Chen, Hui},
  journal={arXiv preprint arXiv:2401.03341},
  year={2024}
}

Acknowlegments

Our codes are influenced by the following repos: VLDB22 and CIKM20