When Should We Admit Generated Nodes? Utility-Guided Few-Shot Learning on Text-Attributed Graphs

Generated nodes offer a promising way to augment limited supervision in text-attributed graphs but their usefulness depends on whether they improve learning rather than merely appear plausible. We study this problem as selective synthetic-node admission where accepting a candidate determines both its use as supervision and its participation in message passing. We introduce CSNA which ranks candidates using semantic and structural signals learned from validation-derived utility, retains a selected top fraction and removes rejected nodes together with their attachment edges. Across Cora and PubMed, CSNA outperforms admit-all across all 1/2/5/10-shot settings, with one-shot accuracy gains of 6.38 and 5.33 percentage points, respectively. Under 40% synthetic-label corruption, the Cora one-shot gain reaches 17.80 points. Together, these results show that the value of synthetic augmentation lies not only in what is generated, but in what is ultimately admitted.