Graph-Attention Diffusion for Enhanced Multivariate Time-Series Anomaly Detection

Abstract

Multivariate time-series anomaly detection is a complex task that requires capturing temporal and spatial correlations. Recently, among the unsupervised methods, diffusion models have attracted increased attention among researchers for addressing this particular task. However, spatial information often remains underutilized or overlooked in existing models. In this article, we propose a novel reconstruction-based approach that enhances normal pattern learning through data masking and leverages diffusion models to capture both temporal and spatial interrelations via graph-attention layers. To address the problem of overgeneralization, where anomalous points are reconstructed too well, potentially abnormal points are initially masked based on the reconstruction error produced by the autoencoder. The masked time-series data is then corrupted by noise and reconstructed back by the diffusion model that removes noise in a step-by-step manner. Evaluation on four datasets from various sources demonstrates the effectiveness of our approach, achieving an average $F1$ -score of 96.51%, outperforming many existing baselines. The ablation study estimates the contribution of each of the key components of the model to the score.