Spiking Graph Predictive Coding for Reliable OOD Generalization

TL;DR

SpIking GrapH predicTive coding (SIGHT), an uncertainty-aware plug-in graph learning module for reliable OOD Generalization, is introduced, which consistently enhances predictive accuracy, uncertainty estimation, and interpretability when integrated with GNNs.

Abstract

Graphs provide a powerful basis for modeling Web-based relational data, with expressive GNNs to support the effective learning in dynamic web environments. However, real-world deployment is hindered by pervasive out-of-distribution (OOD) shifts, where evolving user activity and changing content semantics alter feature distributions and labeling criteria. These shifts often lead to unstable or overconfident predictions, undermining the trustworthiness required for Web4Good applications. Achieving reliable OOD generalization demands principled and interpretable uncertainty estimation; however, existing methods are largely post-hoc, insensitive to distribution shifts, and unable to explain where uncertainty arises especially in high-stakes settings. To address these limitations, we introduce SpIking GrapH predicTive coding (SIGHT), an uncertainty-aware plug-in graph learning module for reliable OOD Generalization. SIGHT performs iterative, error-driven correction over spiking graph states, enabling models to expose internal mismatch signals that reveal where predictions become unreliable. Across multiple graph benchmarks and diverse OOD scenarios, SIGHT consistently enhances predictive accuracy, uncertainty estimation, and interpretability when integrated with GNNs.

Figures (6)

• Figure 1: Overview of the proposed framework SIGHT. In training, the encoded node representations are learned by aggregating neighbor information through GNNs. Before propagation, predictions are refined via an iterative predictive coding (PC) unit over LIF neurons. In testing, the model will be tested on datasets with covariate shift, where the input feature distribution changes while the underlying labels remain stable (i.e., $P^{\text{train}}(X)\neq P^{\text{test}}(X)$ but $P^{\text{train}}(Y|X)=P^{\text{test}}(Y|X)$), and with concept shift, where the meaning of labels or the feature–label relationship itself changes over time (i.e., $P^{\text{train}}(Y|X)\neq P^{\text{test}}(Y|X)$).
• Figure 2: Comparison of label distribution between training and testing sets on Twitch and CBAS.
• Figure 3: Comparison of feature distribution between training and testing sets on Cora, Citeseer, and Pubmed.
• Figure 4: Correlation between PC error and confidence scores on the Cora, Citeseer, and Pubmed OOD sets. Higher PC errors consistently correspond to lower confidence, showing that PC residuals serve as an intrinsic measure of uncertainty.
• Figure 5: Correlation between PC error and confidence scores on the Cora, Citeseer, and Pubmed ID test sets. Higher PC errors consistently correspond to lower confidence, showing that PC residuals serve as an intrinsic measure of uncertainty.