Mining Fluctuation Propagation Graph Among Time Series with Active Learning

Faults are inevitable in a complex online service system. Compared with the textual incident records, the knowledge graph provides an abstract and formal representation for the empirical knowledge of how fluctuations, especially faults, propagate. Recent works utilize causality discovery tools to construct the graph for automatic troubleshooting but neglect its correctness. In this work, we focus on structure discovery of the fluctuation propagation graph among time series. We conduct an empirical study and find that the existing methods either miss a large proportion of relations or discover almost a complete graph. Thus, we propose a relation recommendation framework named FPG-Miner based on active learning. The experiment shows that operators' feedback can make a mining method to recommend the correct relations earlier, accelerating the trustworthy application of intelligent algorithms like automatic troubleshooting. Moreover, we propose a novel classification-based approach named CAR to speed up relation discovery. For example, when discovering 20% correct relations, our approach shortens 2.3-42.2% of the verification quota compared with the baseline approaches.