Correlation-Aware Spatial-Temporal Graph Learning for Multivariate Time-Series Anomaly Detection.

Area of Science:

• Data Science
• Artificial Intelligence
• Machine Learning

Background:

• Multivariate time-series anomaly detection is crucial for various industries.
• Existing methods like statistical models and conventional deep learning (DL) models (CNN, LSTM) struggle with nonlinear relations and pairwise correlations.

Purpose of the Study:

• To propose a novel method for multivariate time-series anomaly detection that overcomes limitations of existing approaches.
• To explicitly capture pairwise correlations and spatial-temporal dependencies among variables.

Main Methods:

• Developed correlation-aware spatial-temporal graph learning (CST-GL) method.
• Integrated a multivariate time-series correlation learning (MTCL) module to capture pairwise correlations.
• Employed a spatial-temporal graph neural network (STGNN) with graph convolution network (GCN) for spatial information and dilated convolutions for temporal dependencies.
• Incorporated an unsupervised anomaly scoring component.

Main Results:

• CST-GL effectively detects and diagnoses anomalies in general settings.
• The method demonstrates capability for early anomaly detection across different time delays.
• Experimental results validate the effectiveness of the proposed approach.

Conclusions:

• CST-GL offers an advanced solution for multivariate time-series anomaly detection.
• The method's ability to learn complex correlations and dependencies enhances detection accuracy and timeliness.
• The unsupervised anomaly scoring provides a robust measure of anomaly severity.