Graph neural network solutions for interpretable anomaly detection in IT infrastructure monitoring time series

Anomaly detection is an essential task for many firms and organizations. Identifying unusual patterns in messy multivariate time series can prevent catastrophic events and optimize operations. Traditional statistical methods struggle with high-dimensional data and complex temporal dependencies. In this paper, we propose a novel approach that combines Generative Adversarial Networks (GANs), a reconstruction-based framework, and Graph Neural Networks (GNNs) for effective and interpretable anomaly detection. Our method involves representing multivariate time series as graphs and training two interconnected GANs and an Autoencoder to capture the normal behaviour of the networks. Anomalies are detected by measuring the reconstruction error and the discriminator’s score. Our case study on the signals of the INFN CNAF Tier-1 data centre demonstrates the effectiveness of our approach in terms of robustness and interpretability. We further validate our model on two widely used benchmark datasets for IT infrastructure monitoring, PSM and SMD, obtaining Fβ scores of 0.943 and 0.763, respectively. This work highlights the potential of GNNs in developing interpretable deep learning solutions for real-world applications.