CyCo: A Temporal Cycle Consistency Based Labeling Method for Time Series Data
Fengrui Liu, Haiyang Jiang, Zulong Diao, Yanbiao Li, Gaogang Xie · 2021
Time series analysis can explore the internal relations among data, so as to further realize prediction, pattern discovery and anomaly detection. However, these analyses largely rely on labeled data, which needs to identify particular patterns from entire time series data. Because labeling is costly and domain specific, how to automatically label raw data with predefined patterns becomes a valuable problem. In this paper, we first discuss the particularities that we learned from practical applications of labeling task in terms of inconsistent data length, single exemplar and the needs of explainability. And then, we propose a simple and effective labeling method called CyCo that exploits temporal cycle consistency for local similarity evaluation, and transductive learning for label propagation. It not only supports labeling on arbitrary length data with a single exemplar, but can also pinpoint similar portions between targets and predefined patterns. We implement a prototype of CyCo in a real-world cloud datacenter to label the monitoring data, and experiments demonstrate that our method can automatically label the expected patterns on large scale datasets with fast performance (average 11 ms per sample) and quality results (average 0.96 F1-score).