Reproduce “DACR: Distribution-Augmented Contrastive Reconstruction for Time-Series Anomaly Detection” on OpenResearch

About this paper

Lixu Wang, Shichao Xu, Xinyu Du, Qi Zhu

DACR, a three-stage methodology for time-series anomaly detection, integrates VAE-based distribution augmentation, self-supervised contrastive learning, and an attention-based reconstruction Transformer. The method consistently established state-of-the-art performance, with AUC improvements ranging from 0.4% to 8.2% over existing baselines across nine benchmark datasets.

Key insights

• Augmenting data distributions through VAE-based latent space perturbation is effective in mitigating the 'uniformity problem' of standard contrastive learning, enhancing anomaly detection boundaries.
• Training univariate feature extractors using a distribution-augmented contrastive learning approach leads to more semantically rich and robust representations for time-series data.
• An attention-based Transformer with an input structure that captures both historical and current inter-feature dependencies improves the accuracy and robustness of time-series reconstruction for anomaly detection.

Summary by alphaXiv.