Harnessing Feature Resonance under Arbitrary Target Alignment for Out-of-Distribution Node Detection

TL;DR

This work tackles the challenge of identifying out-of-distribution (OOD) nodes in graphs without relying on labeled in-distribution categories. It introduces Feature Resonance, a dynamics-based phenomenon wherein unknown in-distribution samples exhibit larger representation changes than OOD samples during ID optimization, which can be exploited in a label-agnostic setting. The resulting Resonance-based Separation and Learning (RSL) framework combines a micro-level resonance proxy with a synthetic OOD node strategy (via SGLD) to train an OOD classifier, supported by theoretical error bounds and extensive experiments across 13 real-world graph datasets. Key findings show that RSL achieves state-of-the-art performance in unsupervised OOD node detection, performs well on heterophilic graphs, and remains effective without multi-class labels, offering a practical, scalable solution for robust graph learning. The approach’s reliance on representation dynamics rather than task-specific labels broadens its applicability to diverse domains and real-world OOD detection tasks.

Abstract

Detecting out-of-distribution (OOD) nodes in the graph-based machine-learning field is challenging, particularly when in-distribution (ID) node multi-category labels are unavailable. Thus, we focus on feature space rather than label space and find that, ideally, during the optimization of known ID samples, unknown ID samples undergo more significant representation changes than OOD samples, even if the model is trained to fit random targets, which we called the Feature Resonance phenomenon. The rationale behind it is that even without gold labels, the local manifold may still exhibit smooth resonance. Based on this, we further develop a novel graph OOD framework, dubbed Resonance-based Separation and Learning (RSL), which comprises two core modules: (i) a more practical micro-level proxy of feature resonance that measures the movement of feature vectors in one training step. (ii) integrate with synthetic OOD nodes strategy to train an effective OOD classifier. Theoretically, we derive an error bound showing the superior separability of OOD nodes during the resonance period. Extensive experiments on a total of thirteen real-world graph datasets empirically demonstrate that RSL achieves state-of-the-art performance.

Figures (7)

• Figure 1: (a) We conduct a preliminary study on the changes in ID and OOD node representations during training using a toy dataset. (b) Projections of the representations of ID and OOD nodes onto gradients: $\text{Proj}_{\nabla \ell(\theta_t; \cdot)}\mathbf{x}_i = \frac{\mathbf{x}_i \cdot \nabla \ell(\theta_t; \cdot) }{\parallel \nabla \ell(\theta_t; \cdot) \parallel_2^2}\cdot \nabla \ell(\theta_t; \cdot)$. (c) Schematic of Feature Resonance.
• Figure 2: The performance of using resonance-based score $\tau$ to detect OOD nodes varies with training progress. The higher the AUROC, the better, and the lower the FPR95, the better.
• Figure 3: The performance of using resonance-based score $\tau$ to detect OOD nodes varies with training progress. The higher the AUROC, the better, and the lower the FPR95, the better.
• Figure 4: Performance of detecting OOD nodes with different metrics. $\tau$ represents the resonance-based score, the "Overall Trajectory" represents the total cumulative length of the training trajectory $\hat{F}(\tilde{\mathbf{x}}_i) = \sum_t \tau_i$, and the "Sliding Window" refers to the cumulative $\tau$ within a window of width 10: $\hat{F}_{10}(\tilde{\mathbf{x}}_i) = \sum^t_{t-10} \tau_i$.
• Figure 5: The impact of different sliding window widths on the performance of detecting OOD nodes. When the width is 1, it corresponds to the resonance-based score $\tau$.

Theorems & Definitions (14)

• Definition 1
• Definition 2
• Theorem 1
• Theorem 2
• Definition 3
• Definition 4
• Definition 5
• Theorem 3
• Theorem 4
• Proposition 1