Empowering GNNs for Domain Adaptation via Denoising Target Graph

Graph neural networks (GNNs) have become a dominant paradigm for learning on graph-structured data, with notable models like GraphSAGE, GAT, and GraphConv. These models excel at capturing local node features and structural patterns, leading to success in social network analysis, recommendation systems, and more. However, real-world data often undergo temporal and spatial shifts, causing distributional discrepancies between training and deployment environments. This domain shift hampers GNN performance, especially when structural changes such as edge addition or removal occur. Existing solutions like feature alignment or global distribution matching are insufficient because they overlook the importance of local structural invariants. Recent works have begun exploring structural regularization and pseudo-labeling, but a comprehensive approach that explicitly models structural shifts remains lacking. This context motivates the development of methods that leverage the graph's inherent structure to improve robustness and generalization across domains.

The core challenge is that graph data collected at different times or from different regions exhibit structural discrepancies, such as missing or spurious edges, which degrade GNN performance. Unlike images, where pixel distributions can be aligned, graphs require preserving local connectivity patterns. Structural shifts can lead to nodes being disconnected or falsely connected, disrupting message passing and feature aggregation. The difficulty lies in designing a method that can adapt to these structural variations without requiring extensive labeled data in the target domain. Moreover, existing methods often ignore the structural invariants that could serve as anchors for transfer learning. Addressing this problem is crucial for deploying GNNs in dynamic environments like social media, financial markets, and evolving knowledge graphs.

This paper introduces several innovations: 1) Incorporation of an auxiliary edge denoising task that trains the model to identify real versus fake edges in the target graph, thus learning structural invariants. 2) Theoretical analysis linking this task to the A-distance, showing that it tightens the generalization bound. 3) A joint training framework that combines node classification with edge denoising, enabling the model to learn robust structural features. 4) Empirical validation demonstrating significant accuracy improvements across multiple datasets and transfer scenarios. These innovations collectively enable the model to better handle structural shifts, a key bottleneck in graph domain adaptation.

• �� Construct noisy target graphs by adding sparse random edges to the original target graph. • Use an MLP-based classifier to perform edge classification, distinguishing real edges from added fake edges. • During training, optimize a combined loss: node classification loss on source graph and edge denoising loss on target graph. • The edge denoising loss is computed as the log-likelihood of correctly classifying real versus fake edges. • Theoretical analysis shows that minimizing this loss constrains the A-distance, reducing the classifier disagreement between source and target. • The training loop involves alternating between adding noise, computing edge classification loss, and updating model parameters with gradient descent.

Experiments utilize the Arxiv dataset for temporal domain transfer, and MAG for regional transfer, with additional tests on ogbn-products. The models are based on GraphSAGE with hyperparameters tuned via grid search. Baselines include ERM, DANN, SPECREG, and recent graph-specific methods. Evaluation metrics are accuracy scores over multiple runs. Ablation studies compare different auxiliary tasks, confirming edge denoising's effectiveness. The experiments demonstrate that GraphDeT consistently outperforms baselines, especially in large temporal gaps and significant structural shifts, validating the theoretical insights. The setup emphasizes robustness and scalability, with hyperparameters carefully optimized.

In the temporal transfer from 1950-2007 to 2016-2018, accuracy improved from 7.30% (ERM) to 21.83% (GraphDeT), a 14.54% gain. On MAG regional shifts, average improvements reached 26.75%, with the best results surpassing existing methods by over 20%. Ablation results show that removing the edge denoising task reduces performance, confirming its critical role. Theoretical bounds align with empirical gains, illustrating how structural constraints tighten the generalization gap. The method remains effective across different datasets, transfer directions, and graph sizes, demonstrating its broad applicability.

This approach is suitable for applications requiring robust graph learning across evolving environments, such as social network analysis, recommendation systems, fraud detection, and knowledge graph maintenance. It enables models to adapt to structural changes without extensive retraining or labeling, reducing deployment costs. The method can be integrated into existing GNN pipelines, providing a plug-and-play enhancement for dynamic data scenarios. Long-term, it paves the way for more resilient AI systems capable of continuous learning in real-world, non-stationary environments.

The method's performance hinges on the appropriate setting of noise ratios; excessive noise can impair learning. Computational overhead increases due to the auxiliary edge classification, especially on large graphs. The current validation focuses on node classification; extending to other tasks like link prediction or graph classification requires further research. Sensitivity to hyperparameters and structural noise levels may limit scalability. Future work should explore adaptive noise strategies, more efficient algorithms, and broader task applicability to overcome these limitations.