Strengthening Unsupervised Graph Out-of-Distribution Detection with Structural Resonance

Most existing graph Out-of-distribution (OOD) detection methods rely on in-distribution (ID) category labels to obtain meaningful representations, which limits their applicability in unlabeled settings. Recent representative unsupervised graph OOD detection methods exploit ID–OOD discrepancy in optimization-induced representation dynamics (Resonance), mitigating the dependence on labels. However, these methods still fall short in fully leveraging the intrinsic structural properties of graphs. Moreover, resonance inferred from single-period optimization-induced changes are highly sensitive to stochastic noise and transient gradient effects, which undermines estimation reliability. To address these challenges, we propose Multi-Scale Structural Resonance Estimation (MSRE), which models local structural resonance and performs multi-scale temporal aggregation, and we further provide a consistency-based candidate selection strategy that requires no labeled validation set. Extensive experiments on multiple benchmark graphs demonstrate that the proposed framework consistently outperforms state-of-the-art unsupervised OOD detection methods.