We propose \textbf{CausalTrace}, an unsupervised m...

We propose \textbf{CausalTrace}, an unsupervised multi-scale method based on graph attention networks for detecting trace anomalies and pinpointing root causes. As illustrated in Fig. \ref{fig:model}, \textbf{CausalTrace} consists of two tasks: anomaly detection and root cause analysis. The objective of \textbf{CausalTrace} is to activate root cause analysis whenever a trace anomaly is identified in a microservice system. To comprehensively understand the relationships between system components and facilitate anomaly detection, three graph structures can be employed: causal graphs, topology graphs, and dependency graphs. Through in-depth analysis, we find that causal graphs and topology graphs belong to knowledge graphs, effectively depicting prior information about the system's normal operating state. Meanwhile, the dependency graph serves as a modeling tool primarily used to represent the relationships between components during normal system operation. With the support of knowledge graphs, the dependency graph can deepen the understanding of the system, thereby enhancing anomaly detection capabilities. The abbreviation "DG" for the dependency graph appears relatively late in the manuscript, while the full name has been introduced earlier. It is recommended to define "DG" when the dependency graph is first mentioned and use consistent capitalization for all defined acronyms throughout the manuscript.