Abstract
Time series anomalies can offer information relevant to critical situationsfacing various fields, from finance and aerospace to the IT, security, andmedical domains. However, detecting anomalies in time series data isparticularly challenging due to the vague definition of anomalies and saiddata's frequent lack of labels and highly complex temporal correlations.Current state-of-the-art unsupervised machine learning methods for anomalydetection suffer from scalability and portability issues, and may have highfalse positive rates. In this paper, we propose TadGAN, an unsupervised anomalydetection approach built on Generative Adversarial Networks (GANs). To capturethe temporal correlations of time series distributions, we use LSTM RecurrentNeural Networks as base models for Generators and Critics. TadGAN is trainedwith cycle consistency loss to allow for effective time-series datareconstruction. We further propose several novel methods to computereconstruction errors, as well as different approaches to combinereconstruction errors and Critic outputs to compute anomaly scores. Todemonstrate the performance and generalizability of our approach, we testseveral anomaly scoring techniques and report the best-suited one. We compareour approach to 8 baseline anomaly detection methods on 11 datasets frommultiple reputable sources such as NASA, Yahoo, Numenta, Amazon, and Twitter.The results show that our approach can effectively detect anomalies andoutperform baseline methods in most cases (6 out of 11). Notably, our methodhas the highest averaged F1 score across all the datasets. Our code is opensource and is available as a benchmarking tool.