Periodic Graph-Enhanced Multivariate Time Series Anomaly Detector
This work addresses a domain-specific problem in fields like finance and healthcare, offering an incremental improvement by enhancing graph structures for better anomaly detection.
The paper tackles the problem of multivariate time series anomaly detection by addressing the limitation of static graph structures in representing complex spatio-temporal correlations, proposing a periodic graph-enhanced method that outperforms state-of-the-art models on four real datasets.
Multivariate time series (MTS) anomaly detection commonly encounters in various domains like finance, healthcare, and industrial monitoring. However, existing MTS anomaly detection methods are mostly defined on the static graph structure, which fails to perform an accurate representation of complex spatio-temporal correlations in MTS. To address this issue, this study proposes a Periodic Graph-Enhanced Multivariate Time Series Anomaly Detector (PGMA) with the following two-fold ideas: a) designing a periodic time-slot allocation strategy based Fast Fourier Transform (FFT), which enables the graph structure to reflect dynamic changes in MTS; b) utilizing graph neural network and temporal extension convolution to accurate extract the complex spatio-temporal correlations from the reconstructed periodic graphs. Experiments on four real datasets from real applications demonstrate that the proposed PGMA outperforms state-of-the-art models in MTS anomaly detection.