TL;DR
Graph domain adaptation methods typically try to align all source and target nodes, but structural mismatches often introduce harmful noise. This paper introduces Source Node Influence Pruning (SNIP), a data-centric, plug-and-play framework that identifies and removes structurally incompatible source nodes using multi-dimensional centrality metrics. By refining the source graph before alignment, SNIP effectively eliminates negative transfer and consistently boosts the performance of existing graph adaptation models.
What Problem Does This Paper Address?
Unsupervised Graph Domain Adaptation (UGDA) helps transfer knowledge from a labeled source graph to an unlabeled target graph. Traditional approaches focus heavily on feature-level alignment—using adversarial training or distance metrics to force the source and target representations into a shared latent space.
However, these methods rely on a flawed assumption: that every node in the source graph is useful for alignment. In graph-structured data, a node’s meaning is deeply tied to its local and global topology. Because of structural shifts between different domains, certain source nodes (like extreme hubs or structural outliers) simply do not have semantic counterparts in the target domain. Forcing the model to align these mismatched nodes distorts the latent space, injecting noise and causing negative transfer (where adaptation actually degrades model performance).
The authors shift the research paradigm from how to align to what to align by focusing on data-level refinement.
Methods
The authors propose Source Node Influence Pruning (SNIP), a lightweight, model-agnostic preprocessing pipeline that filters out structural anomalies from the source graph:
- Node Influence Measurement: SNIP characterizes graph topology from four complementary perspectives: Degree, Betweenness, Closeness, and Eigenvector centrality. It calculates the global mean of these metrics in the target domain and measures how much each individual source node deviates from this target baseline.
- Rank-based Normalization: Centrality metrics operate on vastly different scales (e.g., degree is an integer, while betweenness is a tiny fraction). SNIP converts these absolute values into relative percentile ranks within the source graph, mapping them to a uniform $[0, 1]$ scale.
- Structure Pruning: SNIP computes a final influence score for each node via a weighted combination of the normalized ranks. Nodes with the lowest scores—meaning they are the most structurally divergent from the target domain—are pruned according to a predefined ratio ($r$).
The resulting "clean" sub-source graph is then passed to standard graph neural network encoders and alignment components.
Datasets
The framework was evaluated across eight transfer scenarios using five real-world benchmarks:
- Citation Networks (ArnetMiner): ACMv9, Citationv1, and DBLPv7. Nodes represent academic papers, and edges represent citation relationships.
- Social Networks (Twitch): Gamer networks from Germany (DE) and England (EN), where nodes represent users and edges represent social connections.
Key Insights and Ideas
- Data-Centric > Model-Centric: Instead of engineering complex, regularized loss functions to force alignment, simply cleaning the input data by removing structurally incompatible nodes yields substantial performance gains.
- The Pruning Sweet Spot: Randomly dropping nodes consistently hurts performance because it destroys vital structural information. In contrast, targeted structural pruning shows an inverted-U performance curve: pruning too little leaves behind structural noise, while pruning too much discards useful data. The optimal pruning ratio typically lies between 10% and 30%.
- Centrality as a Domain Detector: Rather than using centrality metrics purely to find "important" nodes for feature enhancement, SNIP repurposes them as structural distribution detectors to quantify cross-domain compatibility.
Key Results
SNIP was integrated as a plug-and-play module with two popular backbones, AdaGCN and A2GNN, outperforming competitive baselines:
- Citation Networks: When applied to AdaGCN, SNIP achieved an average improvement of 3.01% in Micro-F1 and 4.62% in Macro-F1 across six transfer scenarios. It achieved a massive 7.79% Macro-F1 boost in the challenging Citationv1 $\rightarrow$ ACMv9 scenario.
- Social Networks: On the dense Twitch datasets, SNIP improved A2GNN by 1.17% in Macro-F1 and 1.22% in Micro-F1 on average.
- Ablation Performance: Ablation studies revealed that the Rank-based Normalization mechanism was the most critical component; removing it caused the most severe performance drops, validating its necessity in balancing disparate structural scales.
Limitations and Future Work
The authors note that computing global network metrics, such as betweenness centrality, can become computationally expensive in the worst-case scenario for massive, web-scale graphs.
However, because SNIP's indicator selection is highly flexible and decoupled from the core framework, they suggest that computationally intensive metrics can be seamlessly swapped out for highly parallelizable, linear-time approximations (such as localized PageRank or Average Neighbor Degree) to maintain scalability. Swapping to these alternative configurations still delivers highly competitive performance.
Summary updated