Graph Construction Approaches for Graph Neural Networks-Based Anomaly Detection in Time Series

This article surveys heuristic-based and learning-based approaches for constructing adjacency matrices in Graph Neural Network (GNN)-based anomaly detection for multivariate time series (MTS). Effective adjacency matrix construction is crucial for GNN performance, enabling the model to capture dependencies between time series variables. Heuristic methods, like pairwise correlation, distance-based or dynamic time warping distances, offer simplicity but may struggle with complex relationships. Learning-based approaches, leveraging techniques such as attention mechanisms or graph learning layers, dynamically learn the graph structure from the data, potentially capturing more nuanced dependencies. This survey systematizes existing methods within both paradigms, comparing their strengths, weaknesses, computational complexities. Finally, it identifies open challenges and future research directions, focusing on improving computational time and accuracy of GNN-based anomaly detection using advanced graph construction techniques.