MN-GAT: Incorporating Metric Names into Metric Correlation Graphs for Anomaly Detection

Multivariate time series anomaly detection plays an essential role in various domains. Multiple metrics within the same system are not independent but exhibit specific correlations. Accurately characterizing the correlational relationships is the key issue in conducting multiple metrics analysis. In recent years, graph neural networks have shown outstanding performance in anomaly detection by effectively modeling the correlations among metrics. The challenge lies in accurately representing the interrelationships among metrics using graph structures. Existing approaches depend solely on the temporal patterns of metric time series, neglecting the information embedded in the metric names. In fact, metric names provide concise and meaningful information about the metrics and their relationships of. Therefore, this paper proposes a Metric Name (MN) assisted multi-metric graph attention-based (GAT) anomaly detection model, i.e., MN-GAT. This model automatically constructs a multi-metric correlation graph structure based on the multivariate time series and their names, and performs anomaly detection based on this correlation graph. The attention weights contained in the graph provide explanations for anomalies. Experimental results demonstrate that our method accurately models the interrelationships among metrics and achieves more accurate anomaly detection than baseline methods. Furthermore, it can assist operations personnel in identifying the root causes of failures and troubleshooting.