# Time-Series Anomaly Detection Service at Microsoft

**Authors:** Hansheng Ren, Bixiong Xu, Yujing Wang, Chao Yi, Congrui Huang, Xiaoyu Kou, Tony Xing, Mao Yang, Jie Tong, Qi Zhang  
**arXiv:** [1906.03821](https://arxiv.org/abs/1906.03821)  
**Date:** 2019-06-10  
**Summary page:** https://www.summarizepaper.com/en/paper/1906.03821v1/

## Key points

- Authors: Hansheng Ren, Bixiong Xu, Yujing Wang, Chao Yi, Congrui Huang, Xiaoyu Kou, Tony Xing, Mao Yang, Jie Tong, and Qi Zhang
- Critical need for large companies to monitor metrics like Page Views and Revenue in real-time
- Development of an innovative anomaly detection service by the team at Microsoft
- Detailed pipeline with three key modules: data ingestion, experimentation platform, and online compute
- Proposal of a novel algorithm combining Spectral Residual (SR) with Convolutional Neural Network (CNN) techniques
- Pioneering use of SR model from visual saliency detection domain in time-series anomaly detection
- Aim to enhance SR model performance significantly by integrating it with CNN
- Experimental results showing outperformance of state-of-the-art baselines on public datasets and real-world Microsoft production data

## Summary

In their paper titled "Time-Series Anomaly Detection Service at Microsoft," authors Hansheng Ren, Bixiong Xu, Yujing Wang, Chao Yi, Congrui Huang, Xiaoyu Kou, Tony Xing, Mao Yang, Jie Tong, and Qi Zhang address the critical need for large companies to monitor various metrics such as Page Views and Revenue in real-time. To assist customers in continuously monitoring time-series data and alerting them to potential incidents promptly, the team at Microsoft has developed an innovative anomaly detection service. The paper introduces a detailed pipeline and algorithm designed to be accurate, efficient, and applicable across different scenarios. The pipeline comprises three key modules: data ingestion, experimentation platform, and online compute. To tackle the challenge of time-series anomaly detection effectively,<Organization> proposes a novel algorithm that combines Spectral Residual (SR) with Convolutional Neural Network (CNN) techniques. A notable aspect of their work is the pioneering use of the SR model from the visual saliency detection domain in time-series anomaly detection. By integrating SR with CNN in a unique way,<Organization> aim to enhance the performance of the SR model significantly. The experimental results presented in the paper demonstrate that their approach outperforms state-of-the-art baselines on both public datasets and real-world Microsoft production data.<Organization>'s research not only showcases a cutting-edge solution for time-series anomaly detection but also highlights the importance of leveraging innovative algorithms and techniques to address complex monitoring challenges faced by large enterprises.