[Paper Review] TadGAN: Time Series Anomaly Detection Using Generative Adversarial Networks

Alexander Geiger, Dongyu Liu+3 more|arXiv (Cornell University)|Sep 16, 2020

Anomaly Detection Techniques and Applications4 citations

TL;DR

TadGAN proposes a GAN-based framework for unsupervised time series anomaly detection using LSTM-based generators and critics, enhanced with cycle consistency and reconstruction error computation. It achieves the highest average F1 score across 11 datasets by combining Critic outputs with window-based reconstruction errors via multiplication, significantly reducing false positives.

ABSTRACT

Time series anomalies can offer information relevant to critical situations facing various fields, from finance and aerospace to the IT, security, and medical domains. However, detecting anomalies in time series data is particularly challenging due to the vague definition of anomalies and said data's frequent lack of labels and highly complex temporal correlations. Current state-of-the-art unsupervised machine learning methods for anomaly detection suffer from scalability and portability issues, and may have high false positive rates. 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. To demonstrate the performance and generalizability of our approach, we test several anomaly scoring techniques and report the best-suited one. We compare our approach to 8 baseline anomaly detection methods on 11 datasets from multiple reputable sources such as NASA, Yahoo, Numenta, Amazon, and Twitter. The results show that our approach can effectively detect anomalies and outperform baseline methods in most cases (6 out of 11). Notably, our method has the highest averaged F1 score across all the datasets. Our code is open source and is available as a benchmarking tool.

Motivation & Objective

• To address the challenge of unsupervised time series anomaly detection in the absence of labeled data and complex temporal correlations.
• To overcome limitations of existing deep learning methods, such as high false positive rates and poor generalization, by integrating GANs with reconstruction-based anomaly scoring.
• To improve anomaly detection robustness by combining Critic outputs with multiple reconstruction error types (point-wise, window-based) using multiplication.
• To establish a modular, open-source benchmarking framework for evaluating time series anomaly detection methods.
• To demonstrate that post-processing fusion of reconstruction and adversarial signals yields superior performance over relying on either component alone.

Proposed method

• Employs LSTM-based generators and critics to model temporal dependencies in time series data.
• Implements cycle consistency loss to improve reconstruction fidelity and ensure temporal coherence.
• Computes reconstruction errors using both point-wise and window-based methods, with Dynamic Time Warping (DTW) showing superior performance.
• Combines Critic scores and reconstruction errors using multiplication, which amplifies high anomaly scores and improves stability.
• Uses a modular design allowing integration with any reconstruction-based time series model.
• Applies a multi-stage anomaly scoring pipeline: reconstruct input via generator, compute error with Critic, and fuse scores for final anomaly detection.

Experimental results

Research questions

• RQ1 Can a GAN-based framework effectively detect both point and collective anomalies in unlabeled time series data?
• RQ2 How does combining reconstruction errors with Critic outputs improve anomaly detection performance compared to using either alone?
• RQ3 Which reconstruction error computation method—point-wise or window-based—yields more robust and accurate anomaly detection?
• RQ4 Does multiplication of reconstruction and Critic scores outperform convex combination in terms of F1 score and stability?
• RQ5 Can TadGAN generalize across diverse real-world datasets from domains such as aerospace, finance, and cloud computing?

Key findings

• TadGAN achieved the highest average F1 score (0.629) across all 11 datasets, outperforming all 8 baseline methods.
• The combination of Critic scores and window-based reconstruction errors (specifically DTW) yielded the best performance, with an F1 score of 0.629.
• Multiplication of reconstruction and Critic scores consistently outperformed convex combination, achieving higher F1 scores and lower standard deviations.
• Using only the Critic score led to high variance and poor performance on certain datasets (e.g., A4, A3), with an average F1 of 0.393.
• The window-based reconstruction error (DTW) alone achieved an F1 of 0.514 on the MSL dataset, while the fused score reached 0.623.
• TadGAN outperformed baselines on 6 out of 11 datasets, demonstrating strong generalization and robustness.