- Rating
- 1
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This work presents CSNA, a utility-guided framework for selecting synthetic nodes to boost few-shot node classification on text-attributed graphs. Existing methods add generated nodes to mitigate label shortage, yet these nodes may carry wrong labels, redundant content or spurious links and hurt model performance. CSNA extracts semantic and structural features for each candidate synthetic node, evaluates node utility via leave-one-in validation, and uses logistic regression to score node usefulness. It keeps high-value synthetic nodes and deletes low-quality ones together with their edges. Different from common loss reweighting, this hard-selection mechanism impacts both supervised training and graph message passing.
Critical strengths
-
The paper addresses an important and practically relevant problem in graph data augmentation: determining whether a generated node should be included in the training graph. The distinction between loss reweighting and structural admission is well motivated, since generated nodes influence both supervision and message passing. This perspective provides a useful extension to existing synthetic-node generation frameworks.
-
The authors acknowledge several important limitations, including the dependence on additional labeled validation data, computational overhead, moderate utility-prediction generalization, and potential class imbalance introduced by global selection. This discussion helps clarify the conditions under which the proposed method is most effective.
Critical weaknesses
-
The presentation of the proposed method could be improved in terms of clarity and readability. The authors place the complete pseudocode of Algorithm 1 in Appendix C, while repeatedly referring to specific line numbers of this algorithm throughout Section 4.
-
Although the student model is trained using only a few labeled support nodes per class, the proposed framework additionally relies on 500 labeled validation examples for utility estimation, admission-policy selection, and early stopping. Consequently, the setting is better characterized as validation-assisted few-shot learning rather than learning under a strictly limited annotation budget, as the authors acknowledge. This distinction is particularly important because the admission mechanism directly depends on the additional labeled validation data.
I consider this one of the most critical concerns of the paper, as it fundamentally raises questions about the practical motivation and necessity of the proposed framework. If 500 additional labeled nodes are already available, a natural alternative would be to directly incorporate these labels into supervised training or leverage them through existing semi-supervised learning approaches, rather than using them to guide synthetic-node admission. However, the paper does not provide comparisons against such alternatives under the same total annotation budget. Therefore, it remains unclear whether the proposed framework offers meaningful advantages over simply utilizing the available real labeled data. Without such comparisons, the practical value of the proposed method in genuinely label-scarce scenarios is not sufficiently justified.
-
The experimental results do not convincingly demonstrate that the learned utility-guided ranking is essential to the observed improvements. In the matched feature ablations, removing individual features frequently produces comparable or even better performance. Some simpler variants, such as structure-only and utility-only selection, also outperform the full model in certain settings.
-
Insufficient comparisons with existing methods. The experimental evaluation lacks comparisons with existing few-shot node classification methods and relevant graph augmentation approaches. Although the authors include several baselines in Section 7.2, these mainly consist of simple selection strategies, robust loss functions, and the admit-all baseline. Comparisons with established methods from the literature are largely missing, making it difficult to assess the competitiveness of CSNA against existing approaches. Moreover, given the substantial amount of additional labeled validation data used by CSNA, it is important to ensure that competing methods are evaluated under a consistent annotation budget and experimental protocol. The authors should include stronger baselines from existing few-shot graph learning and synthetic graph augmentation methods to better demonstrate the effectiveness of the proposed approach.
-
The reported held-out utility-prediction AUROC is only , suggesting moderate predictive performance and substantial variation across random seeds.
-
A central motivation of the paper is that rejecting synthetic nodes should remove both their supervision and their influence on message passing. However, the current experimental comparisons do not fully isolate these two effects. In particular, the soft-weighting baselines involve different weighting or calibration mechanisms, making it difficult to attribute performance differences exclusively to structural removal. A more convincing experiment would use the same candidate ranking and supervision weights while varying only whether rejected nodes remain in the graph. Such a controlled comparison would directly validate the claimed advantage of hard admission over loss-only filtering.
-
The proposed method introduces several design choices and hyperparameters whose effects are not sufficiently investigated. For example, in Section 4.2 (Equation 4), the authors estimate candidate utility by averaging the validation accuracy differences over three random initializations. Although this is intended to reduce randomness, the choice of exactly three initializations is not justified, and it remains unclear whether the resulting utility estimates and candidate rankings are sufficiently stable. Furthermore, the framework involves multiple hyperparameters, including the retention ratio, attachment budget, neighborhood size, and parameters associated with utility estimation.
-
The effectiveness of synthetic-node admission is inherently related to the quality and characteristics of the generated candidates. Different generation strategies may produce synthetic nodes with varying semantic consistency, diversity, label reliability, and structural compatibility, which can directly affect both the learned utility scores and the final admission decisions.
Questions for authors
See weaknesses section for detailed questions.