Boundary Embedding Shaping with Adaptive Contrastive Learning for Graph Structural Disentanglement

Abstract: Graph neural networks (GNNs) excel at aggregating neighbor information for classification, yet their performance is hindered by graph structural entanglement, where spurious correlations from semantically irrelevant neighbors contaminate node embeddings. This challenge is most acute for nodes near class boundaries in the embedding space, where amplified structural noise blurs decision boundaries and destabilizes predictions. Existing robust GNN methods largely treat all nodes uniformly, ignoring boundary vulnerabilities. In this paper, to improve classification performance, we tackle graph structural disentanglement by identifying boundary-region entanglement as the primary bottleneck and propose Boundary Embedding Shaping (BES), an adaptive contrastive learning GNN plug-in module that selectively suppresses spurious structural noise at decision boundaries with minimal model parameter perturbation. Extensive experiments demonstrate that BES consistently improves boundary discrimination and outperforms existing leading methods. Notably, BES boosts GCN performance by an average of 3.3% in node classification (up to 5.0% on WikiCS) and achieves superior accuracy in link prediction.

• The paper presents BES, a boundary-centric plug-in leveraging adaptive contrastive learning to disentangle class-relevant features from structural noise in GNNs.
• It employs hard example mining and a dynamic attention mechanism to align boundary node embeddings with class centroids, resulting in up to 5.0% accuracy improvement on WikiCS.
• Theoretical analyses confirm that minimizing contrastive loss enhances invariant information extraction while reducing variant noise, ensuring robust and efficient node classification.

Boundary Embedding Shaping with Adaptive Contrastive Learning for Graph Structural Disentanglement

Problem Formulation and Motivation

Graph neural networks (GNNs) have established themselves as the dominant paradigm in graph representation learning, yet they are fundamentally susceptible to structural entanglement, where node embeddings are corrupted by spurious correlations from semantically irrelevant neighbors. This phenomenon is particularly detrimental for nodes proximate to class boundaries, leading to indistinct decision regions and compromised classification accuracy. Prevailing robust GNNs largely address structural noise at a global level, neglecting the disproportionate vulnerability of boundary nodes. The paper introduces a principled boundary-centric approach for structural disentanglement, positing that robust node-level discrimination hinges on precisely targeting entanglement at decision boundaries.

Theoretical Foundations

The authors develop a latent variable model (LVM) that encapsulates the graph data generative process, decomposing latent variables into invariant and variant components. Invariant variables ($Z_{\text{inv}}$) encode class-relevant information resilient to neighborhood variance, while variant variables ($Z_{\text{var}}$) capture dynamic, often noisy, structure. The theoretical analysis proves block-identifiability: minimizing a contrastive loss yields embeddings containing exclusively the information from invariant factors, isolating class-relevant semantics and discarding topological noise.

Key results include:

• Cor. 3.1: Minimizing the contrastive loss provably maximizes mutual information between embeddings and $Z_{\text{inv}}$, while minimizing information about $Z_{\text{var}}$.
• Prop. 3.1: The disentanglement error bound is inversely proportional to boundary margin; thus, increasing boundary separability directly reduces error.

This theoretical underpinning justifies a selective, region-focused loss operationalized via adaptive contrastive learning, targeting hard boundary nodes with maximal structural diversity.

Methodology: Boundary Embedding Shaping (BES)

BES is proposed as a general-purpose GNN plug-in, architected to suppress structural noise at boundaries with minimal parameter perturbation. The pipeline includes:

1. Boundary Node Identification: Hard example mining is applied to detect nodes with maximum neighbor discrepancy, operationalized via a normalized shift score in embedding space and margin-based boundary region detection.
2. Contrastive Objective: The gravity loss systematically aligns boundary node embeddings to class centroids and repels them from competing centers, instantiated via pairwise Euclidean terms. The center-based approximation achieves gradient equivalence, mitigating prohibitive $O(N^2)$ complexity.
3. Adaptive Learning Rate: The update magnitude for the boundary attention layer is dynamically scaled to prevent embedding overshooting, directly reflecting the geometric margin-error relationship. Parameter updates are accepted only when they minimally and efficiently increase boundary separation.

Attention layers are applied solely to boundary nodes, with iterative optimization and residual connections to counteract feature-level coupling. The decoder (e.g., Chebyshev convolution) then performs classification with refined, noise-suppressed embeddings.

Experimental Evaluation

BES is evaluated across homogeneous, heterogeneous, heterophilic, and OGB benchmarks for node classification and link prediction. Strong numerical results are reported:

• Node classification: BES achieves up to 5.0% accuracy improvement on WikiCS and boosts GCN by an average of 3.3%.
• Link prediction: BES consistently surpasses baselines, especially for boundary-spanning edges.

Detailed ablation studies demonstrate that BES’s boundary-focused refinement yields significant improvements from early layers, with diminishing returns upon sequential stacking. Empirical evidence confirms that selective boundary shaping clarifies decision regions without destabilizing non-boundary nodes. Efficiency analysis shows the center-based approximation maintains identifiability properties while reducing convergence time. Further, hyperparameter sensitivity experiments validate robustness to $T$ (temperature), $\delta$ (boundary slab width), and update scale.

Related Work Contrast

The paper synthesizes prior approaches—feature-level disentanglement, structure-level discrimination, class-guided contrastive paradigms, and causal latent block identifiability—and demonstrates that explicit boundary shaping, with controlled parameter updates and adaptive contrastive loss, addresses fundamental limitations of both classical and recent GNN architectures. BES thus offers a principled separation between feature coupling and structural noise, producing robust, expressive representations for downstream tasks.

Practical and Theoretical Implications

BES’s boundary-centric architectural decoupling enables GNNs to attain structure-level disentanglement, enhancing discriminability, robustness, and transferability in graph learning. The theoretical guarantees support the extension of BES to low-label and self-supervised settings, broadening applicability to large-scale graphs with limited annotation. The approach directly addresses cross-node interference and message-passing contamination prevalent in real-world topologies, impacting domains from social network analysis to molecular graph modeling.

Conclusion

The presented work systematically isolates and refines embeddings of boundary nodes, immunizing graph models against structural entanglement and spurious correlations. Theoretical analysis and empirical validation converge to show that adaptive contrastive learning for boundary shaping substantially improves node and edge-level representation quality. The framework promises advancements in graph learning for both supervised and semi-supervised scenarios, with practical scalability and theoretical rigor (2606.20283).