Prospective Multi-Graph Cohesion for Multivariate Time Series Anomaly Detection

2025-09-21

AI Summary

Existing multivariate time series anomaly detection methods predominantly rely on a single graph structure, limiting their capacity to model complex, heterogeneous spatiotemporal dependencies among variables. To address this, we propose PMGC—a multi-graph collaborative anomaly detection framework—that jointly models long-term static prior graphs and short-term dynamic graphs. PMGC introduces a forward-looking graph construction strategy to capture real-time evolution of variable correlations and designs a graph condensation loss function to jointly optimize spatiotemporal dependencies. Notably, PMGC is the first framework to synergistically integrate static knowledge guidance with instance-level dynamic dependency learning within a graph neural network architecture. Extensive experiments on multiple real-world datasets demonstrate that PMGC significantly outperforms state-of-the-art methods, particularly exhibiting superior robustness and detection accuracy in high-dimensional and non-stationary scenarios.

Technology Category

• Machine Learning: Graph-based Machine Learning
• Data Mining & Knowledge Management: Anomaly/Outlier Detection
• Reasoning under Uncertainty: Graphical Models

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Web Mining and Content Analysis: Models for Web evolution
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Anomaly detection in high-dimensional time series data is pivotal for numerous industrial applications. Recent advances in multivariate time series anomaly detection (TSAD) have increasingly leveraged graph structures to model inter-variable relationships, typically employing Graph Neural Networks (GNNs). Despite their promising results, existing methods often rely on a single graph representation, which are insufficient for capturing the complex, diverse relationships inherent in multivariate time series. To address this, we propose the Prospective Multi-Graph Cohesion (PMGC) framework for multivariate TSAD. PMGC exploits spatial correlations by integrating a long-term static graph with a series of short-term instance-wise dynamic graphs, regulated through a graph cohesion loss function. Our theoretical analysis shows that this loss function promotes diversity among dynamic graphs while aligning them with the stable long-term relationships encapsulated by the static graph. Additionally, we introduce a "prospective graphing" strategy to mitigate the limitations of traditional forecasting-based TSAD methods, which often struggle with unpredictable future variations. This strategy allows the model to accurately reflect concurrent inter-series relationships under normal conditions, thereby enhancing anomaly detection efficacy. Empirical evaluations on real-world datasets demonstrate the superior performance of our method compared to existing TSAD techniques.

Problem

• Capturing complex diverse relationships in multivariate time series data
• Overcoming limitations of single graph representations for anomaly detection
• Addressing unpredictable future variations in forecasting-based TSAD methods

Innovation

• Integrates long-term static and short-term dynamic graphs
• Uses graph cohesion loss to align diverse graph relationships
• Employs prospective graphing to capture concurrent inter-series relationships