Reproduce “Time-Series Anomaly Detection Service at Microsoft” on OpenResearch

About this paper

Hansheng Ren, Bixiong Xu, Yujing Wang, Chao Yi, Congrui Huang, Xiaoyu Kou, Tony Xing, Mao Yang and 2 more

Microsoft researchers introduce a time-series anomaly detection service that adapts Spectral Residual (SR) from visual saliency detection, enhancing it with a Convolutional Neural Network (SR-CNN) trained on synthetic anomalies. This unsupervised approach achieves superior accuracy and efficiency, generalizing effectively across diverse time-series patterns for real-time monitoring of millions of metrics.

Key insights

• Principles from visual saliency detection, specifically Spectral Residual, can be effectively transferred to time-series anomaly detection, providing an efficient and unsupervised method.
• Training a CNN discriminator on synthetic anomalies derived from real-world, unlabeled data enables significant accuracy improvements over simple thresholding without requiring manual labels.
• The output of the Spectral Residual model can serve as a powerful feature for enhancing even state-of-the-art supervised Deep Neural Network (DNN) anomaly detection models.

Summary by alphaXiv.