Dual Alignment for Covariate Shift: A Principled Framework for Graph Domain Adaptation

Hongwei Wen, Can Zhang, Hanyuan Hang, Minglong Lei

Keywords: Graph domain adaptation, covariate shift, dual alignment, graph neural network

Abstract: \textit{Graph Domain Adaptation} (\textit{GDA}) is fundamentally challenged by \textit{Covariate Shift} (\textit{CS}), a pervasive discrepancy between source and target graph distributions. We decompose CS into two complementary components: \textit{Feature Shift} (\textit{FS}), arising from mismatched node feature distributions, and \textit{Feature-Conditional Structure Shift} (\textit{FCSS}), reflecting structural variations conditioned on features. Both FS and FCSS distort \textit{Graph Neural Network} (\textit{GNN}) representations, thereby hindering reliable cross-domain transfer. To overcome these issues, we propose \textit{Dual Alignment for Covariate Shift} (\textit{DACS}), a framework that jointly addresses FS and FCSS through adversarial feature alignment for domain-invariant embeddings and adaptive reweighting to enforce structural consistency. Extensive experiments on benchmark datasets demonstrate that DACS effectively bridges domain gaps and consistently outperforms state-of-the-art baselines, highlighting its strong cross-domain generalization.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 13110