Cross-Dataset Time Series Anomaly Detection for Cloud Systems

Explore a 21-minute conference talk from USENIX ATC '19 that delves into cross-dataset time series anomaly detection for cloud systems. Learn about ATAD (Active Transfer Anomaly Detection), an innovative approach that combines transfer learning and active learning techniques to detect anomalies in unlabeled datasets. Discover how this method addresses the challenges of obtaining sufficient labeled data in cloud monitoring environments, where the velocity, volume, and diversity of data make traditional anomaly detection difficult. Understand how ATAD transfers knowledge from existing labeled datasets to new unlabeled ones, and how it determines informative labels for a small portion of unlabeled samples. Gain insights into the effectiveness of this approach, which achieves significant performance improvements while only requiring labeling of 1%-5% of unlabeled data. This talk, presented by researchers from Microsoft Research, Nanjing University, and The University of Newcastle, offers valuable knowledge for maintaining high service availability in cloud computing platforms.

USENIX ATC '19 - Cross-dataset Time Series Anomaly Detection for Cloud Systems