Proactive structural stabilization for robust learning on heterogeneous graphs.

Heterogeneous graph neural networks capture rich heterogeneous semantics through meta-path modeling, but such semantic propagation may also amplify the prediction instability under structural perturbations. Existing robustness methods are often built upon specific assumptions about perturbation patterns, such as identifying and removing unreliable edges, which limits their generalization when structural perturbations deviate from these assumptions. To address this limitation, we propose a proactive structural stabilization framework for robust learning on heterogeneous graphs, termed Proactive STructural stAbilization for roBust LEarning (PSTABLE). During vulnerability assessment, PSTABLE identifies vulnerable nodes by jointly evaluating structural influence, predictive instability, and semantic participation from complementary perspectives. Based on this analysis, PSTABLE introduces learnable auxiliary nodes along key meta-paths to locally strengthen structural support around vulnerable regions. With model parameters fixed, PSTABLE optimizes auxiliary node representations using the prediction confidence of vulnerable nodes as feedback, thereby enhancing local semantic support while preserving semantic integrity. Extensive experiments under both poisoning and evasion attack settings demonstrate that PSTABLE consistently improves robustness under diverse structural perturbations while maintaining competitive performance on clean graphs.