LogGC: Novel Approach for Graph-based Log Anomaly Detection

Abstract

The system logs record the state and behavior of the systems at a given moment and represent an essential resource for understanding system issues and their remediation. Log anomaly detection is a crucial component to detect and prevent system faults. AIOps researchers have proposed many successful log anomaly detection methods based on machine learning models. However, these methods do not consider and model the existing dependencies between logs. Recently, there have been some proposed graph-based log anomaly detection methods that model logs and their dependencies into graph structures. This paper briefly reviews the existing graph-based log anomaly detection methods. It proposes LogGC, a novel approach for graph construction from log sequences, and log anomaly detection through graph classification using graph neural networks. We conduct an extensive experimental evaluation of LogGC on three publicly available datasets, showing that the proposed method outperforms all approaches on two benchmark datasets and all graph-based techniques on all three.