Progressive Graph Structure Adjustment for Homophily Shift in Graph Domain Adaptation

Hongwei Wen, Can Zhang, Hanyuan Hang, Minglong Lei

Keywords: Graph domain adaptation, node homophily shift, graph neural network, graph structure adjustment

Abstract: Node homophily shift—the mismatch in the tendency of nodes to have neighbors with the same label between source and target graphs—poses a key challenge for \textit{Graph Domain Adaptation} (\textit{GDA}) without target labels. We introduce \textit{Progressive Structure Adjustment for Homophily Shift} (\textit{PSAHS}), which progressively reduces homophily discrepancies: in the source graph by modifying existing edges and adding new edges for low-homophily nodes, and in the target graph by making analogous adjustments for nodes with consistent label predictions from \textit{Graph Neural Networks} (\textit{GNNs}) and \textit{Multi-Layer Perceptrons} (\textit{MLPs}). After each refinement, GNNs are updated with domain-adversarial training for representation alignment. This interplay of structure adjustment and representation learning mitigates homophily shift, tightens the target error bound, and yields consistent improvements over strong baselines, highlighting the necessity of node homophily alignment for effective cross-graph transfer.

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

Submission Number: 13135