Title:Structural Alignment Improves Graph Test-Time Adaptation

Abstract:Graph-based learning has achieved remarkable success in domains ranging from recommendation to fraud detection and particle physics by effectively capturing underlying interaction patterns. However, it often struggles to generalize when distribution shifts occur, particularly those involving changes in network connectivity or interaction patterns. Existing approaches designed to mitigate such shifts typically require retraining with full access to source data, rendering them infeasible under strict computational or privacy constraints. To address this limitation, we propose a test-time structural alignment (TSA) algorithm for Graph Test-Time Adaptation (GTTA), a novel method that aligns graph structures during inference without revisiting the source domain. Built upon a theoretically grounded treatment of graph data distribution shifts, TSA integrates three key strategies: an uncertainty-aware neighborhood weighting that accommodates structure shifts, an adaptive balancing of self-node and neighborhood-aggregated representations driven by node representations' signal-to-noise ratio, and a decision boundary refinement that corrects remaining label and feature shifts. Extensive experiments on synthetic and real-world datasets demonstrate that TSA can consistently outperform both non-graph TTA methods and state-of-the-art GTTA baselines.