Unsupervised microservice system anomaly detection via contrastive multi-modal representation clustering

Abstract

Anomaly detection in microservice systems is crucial for ensuring system stability and reliability. Existing methods rely solely on a single type of monitoring data (e.g., metrics or logs) with either partial or full labels, which miss a large number of anomalies and is costly to manually tag. Therefore, we propose an unsupervised Microservice system Anomaly Detection method via Contrastive Multi-modal representation Clustering (MAD-CMC) to tackle these issues. MAD-CMC first adopts a hierarchical architecture to simultaneously explore the spatial–temporal correlation of metrics and log context information. Next, to facilitate metrics-logs interaction, MAD-CMC introduces a cross-modal Transformer, which outputs multi-modal representation for clustering. During the clustering process, we design a multi-grained contrastive learning approach. Benefiting from the clustering results, MAD-CMC bring intra-cluster representation closer while pushing inter-cluster representation farther away at both inter- and intra-modality aspect. Considering that normal samples are simpler and far more numerous than abnormal sample, we propose a dynamic weighting formula, and apply it to contrastive loss to improve the model’s discrimination ability. Sufficient experiments on public dataset show that MAD-CMC outperforms state-of-the-art methods.