Title: Refining Pseudo-Labels via Safe Subspaces for Source-Free Graph Domain Adaptation

Abstract: Source-free graph domain adaptation (SF-GDA) focuses on adapting models trained on source data to unlabeled target graphs, particularly when access to the original source graphs is unavailable. A primary challenge in this process is ensuring the reliability of pseudo-labels. Due to shifts in features and topology, predictions derived from the source model may be confidently incorrect, and unselective self-training can exacerbate systematic errors through the mechanism of graph message passing. This study addresses SF-GDA by adopting a selective pseudo-labeling approach. Rather than assuming that pseudo-label noise is globally bounded across the entire target domain, we pinpoint a confidence-consistent safe subspace where noise can be managed under constrained posterior discrepancy. Furthermore, we establish a target-risk decomposition that isolates the fitting error within the safe subspace, the noise associated with selected labels, and the risk arising from uncertain sets.

Building on this theoretical framework, we introduce SafeSubspace Pseudo-Label Refinement (S²PLR), a novel source-free graph adaptation method. S²PLR restricts hard pseudo-label supervision exclusively to target graphs that are validated by both semantic and structural evidence. The method assesses semantic reliability by analyzing source-committee confidence and disagreement, while simultaneously learning a structural representation intrinsic to the target domain through graph contrastive learning. Pseudo-labels are subsequently verified via neighborhood consistency. For the remaining samples deemed uncertain, the framework employs noise-tolerant soft regularization instead of relying on potentially unreliable hard labels. Empirical evaluations on both image and real-world graph benchmarks, subjected to various domain shifts, indicate that S²PLR delivers robust and competitive results across a wide range of source-free transfer scenarios.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC