Reproduce “TadGAN: Time Series Anomaly Detection Using Generative Adversarial Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alexander Geiger, Dongyu Liu, Sarah Alnegheimish, Alfredo Cuesta-Infante, Kalyan Veeramachaneni
Researchers from MIT and Universidad Rey Juan Carlos developed TadGAN, an unsupervised anomaly detection method that leverages a cycle-consistent Generative Adversarial Network for time series data. The model reconstructs time series and combines reconstruction errors with a discriminator's output to identify anomalies, achieving an average F1-score of 0.700 across 11 diverse datasets, establishing itself as a leading method.
Key insights
- Cycle-consistent GANs effectively mitigate the common overfitting problem in reconstruction-based deep learning models for time series anomaly detection.
- Combining reconstruction errors with the Critic's output, particularly through a multiplicative approach, significantly improves anomaly detection performance by leveraging complementary signals.
- Dynamic Time Warping (DTW) proves a more robust reconstruction error measure for time series, outperforming point-wise and area differences by better handling time shifts and subtle deviations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.