Abstract
Graph Domain Adaptation (GDA) is critical for transferring knowledge between networks with different distributions, such as distinct citation databases or social platforms. Traditional methods often fail to account for global structural similarities and suffer from coarse alignment that ignores class-specific geometries. We introduce the Joint Domain Adaptive Graph Convolutional Network (JDA-GCN), which utilizes a Global Consistency Matrix to capture neighborhood sub-structures and a joint alignment strategy to synchronize both domain-wide and class-level distributions. Experimental results on real-world datasets show that JDA-GCN significantly outperforms state-of-the-art baselines, achieving up to a 3.13% accuracy boost on blog interest classification.
Introduction
In cross-network node classification, we frequently encounter a labeled source graph and an unlabeled target graph. Domain Adaptation (DA) seeks to bridge the gap between these distributions. However, applying DA to graphs presents unique challenges: standard Graph Convolutional Networks (GCNs) often ignore "global consistency"—nodes that share similar structural roles but lack direct edges—and adversarial alignment often treats the entire domain as a single entity, leading to "negative transfer" where specific categories remain misaligned.
Methodology
Node Embedding with Global Awareness
To address the limitations of local adjacency, JDA-GCN incorporates a Global Consistency Matrix ($S$). While the standard adjacency matrix ($A$) captures direct links, $S$ uses random walks to identify nodes with similar neighborhood "shapes," regardless of their distance in the graph. By using a propagation matrix $H = A + S$, the model learns from both local connectivity and structural twins.
Joint Structural Alignment
The core of JDA-GCN is its "Joint" alignment strategy. Rather than just aligning the global distribution, the model performs multi-scale alignment:
- Global Alignment: Uses a discriminator to align the overall source and target feature spaces.
- Class-Level Alignment: Employs Gram Matrices to preserve the relative positions of class centers and Maximum Mean Discrepancy (MMD) to minimize distances between nodes of the same category across domains.
Robust Pseudo-Labeling
Since the target domain lacks labels, JDA-GCN uses a non-parametric spherical k-means clustering approach. By initializing target clusters with source domain centers, the model generates reliable pseudo-labels to guide the intra-class alignment process.
Key Results and Discussion
JDA-GCN was evaluated on citation (Citationv1, ACMv9, DBLPv7) and social (Blog1, Blog2) datasets. Key findings include:
- Superior Accuracy: The model achieved an average improvement of 1.65% on citation tasks and 3.13% on blog datasets compared to the best existing baselines.
- Ablation Insights: Removing the Global Consistency Matrix ($S$) led to an 11.30% drop in accuracy on citation tasks, confirming that neighborhood structure is vital for effective transfer.
- Feature Visualization: t-SNE analysis confirmed that JDA-GCN produces tighter clusters and cleaner boundaries, effectively mapping source and target nodes of the same category to the same regions of the latent space.
Conclusion
JDA-GCN addresses the dual challenges of structural blindness and coarse alignment in graph domain adaptation. By integrating global neighborhood consistency and joint class-level alignment, the framework provides a robust solution for cross-network knowledge transfer. Future work will focus on optimizing the computational complexity of the consistency matrix for large-scale graphs and extending the framework to dynamic, time-evolving networks.
Summary updated