Graph domain adaptation improves with combined local and graph experts
Graph Domain Adaptation Does Not End with Representation Learning
Machine Learning
Summary
Adapting knowledge between different graphs is hard when both the connections and features change. The authors show that relying on just one method that uses graph connections might miss important clues. They propose using two separate approaches: one that looks at the graph structure and another that focuses only on node features without connections. Combining these two approaches without mixing their training leads to better performance when predicting in new graphs with different structures.
What this means in practice
- •For data scientists: Improve predictions on networked data when the target data’s structure differs from training data by combining graph-aware and local feature methods.
- •For cybersecurity teams: Enhance detection of threats in new network environments where connection patterns have changed by using complementary graph and feature-based models.
Authors
Ziqian Liu, Yongxue Xu, Enze Zhang, Jiaqi Zhang, Hao Wang, Maolin Wang
Abstract
Graph domain adaptation (GDA) transfers knowledge from a labeled source graph to an unlabeled target graph under shifts in both node attributes and graph structure. Existing methods primarily adapt graph representations through propagation redesign, distribution alignment, or source-to-target transition modeling, but still rely on a single graph-propagating path for target prediction. This leaves open whether an adapted graph representation exhausts the predictive evidence available in the target domain, since the graph-aware expert and graph-free local expert may exhibit different failure modes under topological shifts. To address this limitation, we propose EviGDA, an Evidence-Augmented Graph Domain Adaptation framework that complements graph representation adaptation with a graph-free local expert. The graph-aware expert performs message passing and entropy-aware marginal alignment, while the graph-free local expert learns solely from source node features and labels without graph propagation or target alignment. The two experts are optimized independently and combined only at inference through a task-level constant probability mixture, preserving complementary evidence without joint training, learned routing, or target pseudo-labels. Extensive experiments on ten datasets and 16 transfer tasks show that EviGDA outperforms state-of-the-art baselines.