TadGAN: Time Series Anomaly Detection
Using GANs (2020)
JunPyo Park
Financial Engineering Lab
Department of Industrial Engineering
Alexander Geiger et al.
acXiv
TadGAN
UNIST Financial Engineering Lab. 1
TadGAN - Github
UNIST Financial Engineering Lab. 2
https://github.com/signals-dev/Orion
TadGAN - Github
UNIST Financial Engineering Lab. 3
Abstract
 In this paper, we propose TadGAN, an unsupervised
anomaly detection approach built on Generative
Adversarial Networks (GANs). To capture the temporal
correlations of time series distributions, we use LSTM
Recurrent Neural Networks as base models for
Generators and Critics.
 TadGAN is trained with cycle consistency loss to allow
for effective time-series data reconstruction. We further
propose several novel methods to compute
reconstruction errors, as well as different approaches
to combine reconstruction errors and Critic outputs to
compute anomaly scores.
 Critic ~ Discriminator
(WGAN에서의 Discriminator를 Critic 이라 함)
UNIST Financial Engineering Lab. 4
Time series anomaly detection
UNIST Financial Engineering Lab. 5
Types of anomalies
UNIST Financial Engineering Lab. 6
Collective
Types of anomalies
UNIST Financial Engineering Lab. 7
Types of anomalies
UNIST Financial Engineering Lab. 8
Traditional(simplest) Methods - Thresholding
UNIST Financial Engineering Lab. 9
Statistical Methods - ARIMA
UNIST Financial Engineering Lab. 10
Deep learning-based methods
UNIST Financial Engineering Lab. 11
Performance
UNIST Financial Engineering Lab. 12
Important Question: Do these new, complex
approaches perform better than a simple baseline
statistical method?
Unsupervised Time Series Anomaly Detection
UNIST Financial Engineering Lab. 13
M개
T기간
Problem Setting
Difficulties
- Do not have previously identified “known anomalies”
- Non availability of “normal baselines”
- Not all detected anomalies are problematic (e.g., Regime change)
- The length of ai is also variable and is not known a priori.
- Evaluation Problem: How to evaluate?
Unsupervised Time Series Anomaly Detection
UNIST Financial Engineering Lab. 14
Methodology: Proximity-based methods (KNN, LOF)
Unsupervised Time Series Anomaly Detection
UNIST Financial Engineering Lab. 15
Methodology: Prediction-based methods (ARIMA, LSTM, RNNs…)
Unsupervised Time Series Anomaly Detection
UNIST Financial Engineering Lab. 16
Methodology: Reconstruction-based methods (AE, VAE, GANs…)
DCGAN(Simplest)
UNIST Financial Engineering Lab. 17
WGAN
UNIST Financial Engineering Lab. 18
WGAN
UNIST Financial Engineering Lab. 19
TAnoGAN
UNIST Financial Engineering Lab. 20
TAnoGAN
UNIST Financial Engineering Lab. 21
TAnoGAN
UNIST Financial Engineering Lab. 22
TAnoGAN
UNIST Financial Engineering Lab. 23
TAnoGAN
UNIST Financial Engineering Lab. 24
MADGAN
UNIST Financial Engineering Lab. 25
MADGAN
UNIST Financial Engineering Lab. 26
TadGAN – What is difference?
UNIST Financial Engineering Lab. 27
- GAN-reconstruction based anomaly detection method for time series data.
- Introduce a cycle-consistent GAN architecture for time-series-to-time-series
mapping
- Identify two time series similarity measures suitable for evaluating the contextual
similarity between original and GAN-reconstructed sequences.
- Conduct extensive evaluation
- Develop a benchmarking system
WGAN, BiGAN, CycleGAN, Timeseries GAN 등을 참고하여
개선
BiGAN
UNIST Financial Engineering Lab. 28
CycleGAN
UNIST Financial Engineering Lab. 29
CycleGAN
UNIST Financial Engineering Lab. 30
CycleGAN
UNIST Financial Engineering Lab. 31
TadGAN – What is difference?
UNIST Financial Engineering Lab. 32
TadGAN – What is difference?
UNIST Financial Engineering Lab. 33
Cycle-consistency Loss
Wasserstein loss
Full Objective
TadGAN – What is difference?
UNIST Financial Engineering Lab. 34
Input: Sliding Window
Training
Reconstruction
TadGAN – What is difference?
UNIST Financial Engineering Lab. 35
Reconstruct the whole time series
Calculate Reconstruction Scores: Simple Method
TadGAN – What is difference?
UNIST Financial Engineering Lab. 36
Calculate Reconstruction Scores: DTW
DTW: Dynamic Time Warping
TadGAN – What is difference?
UNIST Financial Engineering Lab. 37
Calculate Critic Scores: Normalized
TadGAN – What is difference?
UNIST Financial Engineering Lab. 38
Total score
UNIST Financial Engineering Lab. 39
Thank you for listening!

TadGAN: Time Series Anomaly Detection Using GANs (2020)