Abstract: In the realm of cross-network tasks, graph domain adaptation is an effective tool due to its ability to transfer abundant labels from nodes in the source domain to those in the target domain. Existing adversarial domain adaptation methods mainly focus on domain-wise alignment. These approaches, while effective in mitigating the marginal distribution shift between the two domains, often ignore the integral aspect of structural alignment, potentially leading to negative transfer. To address this issue, we propose a joint adversarial domain adaptive graph convolutional network (JDA-GCN) that is uniquely augmented with structural graph alignment, so as to enhance the efficacy of knowledge transfer. Specifically, we construct a structural graph to delineate the interconnections among nodes within identical categories across the source and target domains. To further refine node representation, we integrate the local consistency matrix with the global consistency matrix, thereby leveraging the learning of the sub-structure similarity of nodes to enable more robust and effective representation of nodes. Empirical evaluation on diverse real-world datasets substantiates the superiority of our proposed method, marking a significant advancement over existing state-of-the-art graph domain adaptation algorithms.
Introduction
Graph Domain Adaptation (GDA) addresses the challenge of transferring knowledge from a label-rich "source" graph to a label-scarce "target" graph. This is particularly relevant in dynamic environments, such as citation networks or social media platforms, where connection patterns and topics shift over time. Traditional methods often struggle with two primary issues: a "local-only" blind spot where models fail to recognize distant but functionally similar nodes, and surface-level alignment that ignores the internal structural relationships of specific categories.
The Joint Domain Adaptive Graph Convolutional Network (JDA-GCN) overcomes these limitations by simultaneously capturing local and global node similarities and aligning the internal "structural fingerprints" of categories across domains.
Methodology
The JDA-GCN architecture employs a three-pronged strategy to ensure robust knowledge transfer:
Holistic Node Embedding
Standard Graph Convolutional Networks (GCNs) typically rely on immediate neighborhood information, missing "global consistency." JDA-GCN integrates a Global Consistency Matrix alongside the standard adjacency matrix. By utilizing random walks to measure sub-structure similarity, the model identifies nodes with similar roles even if they are not directly connected.
Joint Structural Alignment
Beyond standard adversarial training—which attempts to make domains indistinguishable on average—JDA-GCN performs structural alignment. It uses Gram Matrices to model the relationships between class centers and Maximum Mean Discrepancy (MMD) to minimize the distance between these intra-class structures across domains. This ensures that the "skeleton" of the classes is preserved during the transfer process, mitigating the risk of negative transfer.
Robust Pseudo-Labeling
To handle the unlabeled target domain, the model implements a non-parametric spherical k-means strategy. Target labels are initialized using source domain cluster centers and iteratively refined, ensuring that target nodes are aligned with the correct categorical structures.
Experimental Evaluation
Datasets
The model was evaluated on two major benchmarks:
- Citation Datasets: Citationv1, ACMv9, and DBLPv7, representing academic papers and their citation patterns across five categories.
- Blog Datasets: Blog1 and Blog2, representing blogger friendships categorized into six classes.
Key Results
JDA-GCN consistently outperformed state-of-the-art baselines, including UDA-GCN and JHGDA:
- Accuracy: On Citation datasets, JDA-GCN achieved an average improvement of 1.65% over the best baseline. On Blog datasets, the improvement was 3.13%, with a notable 5.00% gain in the Blog1 to Blog2 task.
- Ablation Study: Removing the global consistency component (JDA-GCN¬p) led to an average performance drop of 11.3% on Citation tasks, confirming that accounting for non-local sub-structure similarity is vital for effective graph representation.
Conclusion
JDA-GCN represents a significant advancement in GDA by prioritizing structural alignment over simple distribution matching. While the model shows superior performance, future work may address scalability to massive graphs and adaptation for heterogeneous networks containing multiple node and edge types. Nonetheless, the integration of global consistency and intra-class structural preservation provides a robust framework for cross-network knowledge transfer.
Summary updated