Title: Characterizing Spectral Energy Shifts for Spatio-Temporal Graph Anomaly Detection

Abstract: Graph anomaly detection techniques are primarily designed to identify irregular nodes. Although previous approaches have focused on anomalies characterized by heightened variation in spectral energy distributions, they tend to neglect cases where variation decreases—specifically, camouflaged anomalies that mimic normal behavior. This study demonstrates that such anomalies are prevalent across various datasets and evade detection by current spectral methods. To overcome this gap, we present a node-level spectral energy formulation that integrates seamlessly with message passing mechanisms, thereby facilitating the identification of these hidden anomalies. Leveraging this foundation, we develop an energy-aware graph learning framework that captures spectral shifts via energy-driven message passing, applicable to both static and time-series graphs. Furthermore, our unified architecture adapts to temporal contexts without requiring dedicated sequence modules, allowing for efficient processing even with extended sliding windows. Comprehensive evaluations on large-scale benchmarks confirm the robustness and scalability of our proposed method.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC