acceptodds
Under review as a conference paper at ICLR 2027

BEAST: Boundary-aware Semantic Extrapolation with Debiased Topology for Long-Tailed Text-attributed Graph Learning

Abstract

This paper studies the problem of long-tailed text-attributed graph learning, where severe class imbalance hinders the learning of tail classes due to limited semantic evidence and biased structural contexts. Existing approaches typically augment minority classes to create a balanced graph and train graph neural networks (GNNs) on the augmented data. Despite the progress, their performance is far from satisfactory due to overconcentration among synthesized tail nodes and noisy edges during graph augmentation. Towards this end, we propose a novel framework named Boundary-aware Semantic Extrapolation with Debiased Topology (BEAST) for long-tailed text-attributed graph learning. The core of our BEAST is to generate informative nodes along with reliable edges in augmented graphs from two complementary perspectives, i.e., semantic boundary exploration and structural pattern acquisition. In particular, we first identify semantic class competitors of each tail class by measuring inter-class prototype distance in the latent space. More importantly, we leverage large language models (LLMs) to generate auxiliary subclasses for tail classes, guiding node synthesis toward desired regions while maintaining sufficient separation from competing classes. These auxiliary subclasses would ensure diversity during extrapolation, while boundary-aware filtering guarantees semantic consistency of synthesized nodes. In addition, we sample a class-balanced graph as a calibration set to train a shadow model for predicting connectivity patterns, which are subsequently utilized for debiased topology imputation. To mitigate parent-child competition, we further adjust the prediction logits and optimize the model in a multi-task learning framework. Extensive experiments on multiple long-tailed text-attributed graph benchmarks demonstrate the effectiveness of BEAST against representative baselines. Our code is available at https://anonymous.4open.science/r/BEAST-CC22.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Text-attributed Graph, Long-tailed Learning, Graph Data Augmentation, Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_kMzoOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses node classification on long-tailed text-attributed graphs, where tail classes lack sufficient semantic and structural evidence. BEAST uses LLM-generated auxiliary subclasses to guide semantic augmentation, filters synthetic nodes near class prototype boundaries, and learns their connections through a class-rebalanced shadow model. Hierarchical supervision links the auxiliary subclasses to the original labels, aiming to expand tail-class coverage beyond the observed samples. Experiments on various benchmark report higher accuracy than the evaluated baselines and improvements in most class-balanced metrics.

Critical strengths

S1:The method addresses both the semantic content of synthetic nodes and their graph connections, providing a coordinated approach to tail-class augmentation.

S2:The hierarchical supervision design is well motivated, as it incorporates fine-grained subclass semantics while maintaining consistency with the original classification task.

Critical weaknesses

W1:The ablations do not directly isolate the benefit of boundary-oriented sampling. A comparison with random or non-boundary sampling under the same setting would establish its contribution.

W2:The paper does not report accuracy separately for head and tail classes, making it unclear whether the improvements stem from head classes, tail classes, or both. It’s recommended to report subgroup accuracy and comparing them with the main baselines to verify its effectiveness in improving performance for tail classes.

Questions for authors

Q1:How does boundary-oriented sampling compare with random and non-boundary sampling when the number of accepted synthetic nodes is held fixed?

Q2:What are the separate head and tail-class accuracy, does the improvement stem from head or tail?

Q3:How does performance vary with the number of auxiliary subclasses per tail class?

Q4:While keeping the synthetic nodes and their semantic content unchanged, how much additional benefit does the proposed connection method provide compared to simple strategies?

Q5:Since BEAST utilize a LLM to generate auxiliary nodes, how much additional computational overhead does the BEAST enhancement process introduce?

5 comments
AuthorsOpenReview ↗

Response 1

We are truly grateful for the time you have taken to review our paper, your insightful comments and support. Your positive feedback is incredibly encouraging for us! In the following response, we would like to address your major concern and provide additional clarification.

W1&Q1 Boundary Sampling at a Fixed Augmentation Budget

W1: The ablations do not directly isolate the benefit of boundary-oriented sampling. A comparison with random or non-boundary sampling under the same setting would establish its contribution.

Q1: How does boundary-oriented sampling compare with random and non-boundary sampling when the number of accepted synthetic nodes is held fixed?

R1: Thank you for your comment. We conduct a controlled comparison among random, interior, and boundary-oriented sampling while keeping the number of accepted synthetic nodes fixed. As shown in the Table, BEAST consistently outperforms both alternatives across all five metrics. In particular, compared with random sampling, BEAST improves Macro-F1, balanced accuracy, and tail macro recall by 1.08, 1.32, and 1.33 percentage points, respectively. These results demonstrate that boundary-oriented sampling provides additional benefits beyond simply increasing the number of synthetic nodes, particularly in improving balanced and tail-class recognition.

MethodAccuracy (%)Macro-F1 (%)Balanced accuracy (%)Tail macro recall (%)Tail macro-F1 (%)
Random75.57 ± 0.2573.12 ± 0.1574.67 ± 0.1571.44 ± 0.3972.30 ± 0.33
Interior75.48 ± 0.2573.00 ± 0.3573.74 ± 0.1771.29 ± 0.4972.12 ± 0.53
BEAST76.63 ± 0.6074.20 ± 0.4575.99 ± 0.4872.77 ± 0.4272.44 ± 0.56

W2:The paper does not report accuracy separately for head and tail classes, making it unclear whether the improvements stem from head classes, tail classes, or both. It’s recommended to report subgroup accuracy and comparing them with the main baselines to verify its effectiveness in improving performance for tail classes.

Q2: What are the separate head and tail-class accuracy, does the improvement stem from head or tail?

R2: Thank you for your comment. We thank the reviewer for this valuable suggestion. We report head- and tail-class accuracy separately to identify the sources of performance improvements. As shown in Table, BEAST consistently improves head-class accuracy across all three datasets. More importantly, it substantially improves tail-class accuracy on PubMed and Citeseer by 4.62 and 8.07 percentage points, respectively, despite a slight decrease on Cora. These results demonstrate that BEAST not only enhances head-class recognition but also provides substantial benefits for tail-class recognition.

DatasetMethodHead Acc (%)Tail Acc (%)
CoraSAVE-TAG-S82.7175.51
BEAST84.7274.90
PubMedSAVE-TAG-S81.3859.71
BEAST84.6664.33
CiteseerSAVE-TAG-S61.8761.29
BEAST62.2969.36
AuthorsOpenReview ↗

Response 2

Q3: How does performance vary with the number of auxiliary subclasses per tail class?

R3: Thank you for your comment. We investigate the effect of the number of auxiliary subclasses on the Photo dataset using five random seeds. We choose Photo because it provides a complementary evaluation on an Amazon co-purchase graph and contains 12 categories, including 8 tail classes, offering a suitable setting to examine the impact of auxiliary subclass granularity. As shown in the Table, increasing the number of requested subclasses does not lead to monotonically improved performance. Using four subclasses achieves the highest accuracy of 65.60% and a Macro-F1 of 63.98%. Increasing the number to eight brings negligible improvement in Macro-F1 while slightly reducing accuracy. These results suggest that four auxiliary subclasses offer a favorable balance between semantic diversity and generation effectiveness.

Auxiliary SubclassAccepted nodesAdded undirected edgesAcc (%)Macro-F1 (%)
1853,40064.7862.97
2923,68063.8963.41
4773,08065.6063.98
8632,52064.8163.99

Q4: While keeping the synthetic nodes and their semantic content unchanged, how much additional benefit does the proposed connection method provide compared to simple strategies?

R4: Thank you for your comment. To isolate the contribution of the proposed connection method, we keep the synthetic nodes and their semantic content unchanged while varying the connection strategies. As shown in the Table, our recovery method consistently achieves the highest Macro-F1 across all three datasets, outperforming the strongest alternative by 0.48, 0.67, and 1.50 percentage points on Cora, PubMed, and Citeseer, respectively. It also consistently outperforms the no-edge setting, demonstrating that appropriately recovering connections provides additional benefits beyond semantic augmentation alone. These results validate the effectiveness of the proposed topology recovery strategy.

MethodCoraPubMedCiteseer
Semantic kNN73.8271.5063.60
Random69.0070.6661.96
VGAE70.7171.4263.02
No edges73.2570.8663.24
BEAST74.3072.1765.10

Q5: Since BEAST utilize a LLM to generate auxiliary nodes, how much additional computational overhead does the BEAST enhancement process introduce?

R4: Thank you for your comment. To quantify the additional LLM-related computational overhead, we compare the invocation counts and token consumption of BEAST and SAVE-TAG-S. As shown in the Table, BEAST requires more LLM calls across all three datasets, with increases of 60.0%, 87.5%, and 31.3% on Cora, PubMed, and Citeseer, respectively. Although BEAST requires more LLM calls and incurs comparable token consumption, it achieves better classification performance, demonstrating the effectiveness of its more comprehensive semantic augmentation.

DatasetBEAST LLM callsBEAST input / output tokensSAVE LLM callsSAVE input / output tokens
Cora3239,664 / 19,8972036,331 / 16,711
PubMed1527,639 / 10,355829,041 / 9,541
Citeseer2143,129 / 23,8381633,476 / 17,300

Thanks again for appreciating our work and for your constructive suggestions. Please let us know if you have further questions.

Thank you for the informative rebuttal. My concerns have been adequately addressed, and I have therefore raised my score from 3 to 4.

Reviewer_DuxeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies class-imbalanced ("long-tailed") node classification on text-attributed graphs (TAGs). It identifies two failure modes: tail classes have limited semantic diversity, and the structural context around them is biased toward head classes. BEAST has four stages.

  1. Auxiliary subclass proposal (§3.2)
    • Class prototypes are computed from frozen Sentence-BERT embeddings (Eq. 1).
    • For each tail class, the qq most similar classes are retrieved as competitors (Eq. 2).
    • An LLM proposes fine-grained subclasses, each a keyword plus a description, conditioned on the target and competitor class names (Eq. 3).
    • A proposal is kept if it lies within a distance window around the target prototype and is closer to the target than to the nearest competitor (Eq. 4).
  2. Semantic extrapolation (§3.3)
    • For each tail class, candidate texts are generated and split across its subclasses (Eq. 5).
    • A candidate is kept if its boundary margin satisfies 0<Δ≤τbd0 < Δ ≤ τ_{bd} (Eqs. 6–7) and an LLM critic accepts it.
    • Accepted nodes carry their subclass labels.
  3. Topology imputation (§3.4)
    • A class-rebalanced shadow graph is sampled (Eq. 8). A GNN is trained on it, a subset of edges is removed, and one gradient step is taken (Eq. 9).
    • Confidence-propagation responses (Eq. 10) and pairwise confidence similarity gap features (Eq. 11) feed an MLP edge scorer trained with corresponding training objective (Eq. 12).
    • To give generated nodes a provisional topology, each generated node vv is temporarily linked to the KinitK_{init} original nodes whose text embeddings are most cosine-similar to its own, Nsem(v)N_{sem}(v). These links form εinitε_{init} and the provisional graph GprovG_{prov} (Eq. 13).
    • The frozen shadow model is run on GprovG_{prov}. Each generated node's candidate edges are then scored and kept by a threshold τeτ_e and a top-KedgeK_{edge} rule (Eq. 14).
    • Nodes with fewer than KminK_{min} recovered edges are dropped (Eq. 15).
  4. Hierarchical multi-task training (§3.5)
    • A classifier over C+∣A∣C + |A| labels (the CC original classes plus the auxiliary subclasses AA) is trained with a log-frequency logit adjustment and parent–child "silencing" (Eqs. 16–18).
    • A second, parent-level loss aggregates subclass logits with log-sum-exp (Eqs. 19–20). The two losses are combined (Eq. 21).
    • Inference uses the aggregated parent logits.

Section 3.6 gives two results.

  • Theorem 3.1: The margin changes by at most 2(εh+εμ)2(ε_h + ε_μ) under embedding and prototype perturbations.
  • Theorem 3.2: The expected fraction of false recovered edges is at most 1−τe+εe1 − τ_e + ε_e. This assumes that the edge scores are within εeε_e of the edge probabilities in a "reference random augmented graph".

Setup

  • Six datasets are used: Cora, PubMed, Citeseer, and the Amazon graphs Photo, Computer and Children. The backbone is a GCN, the LLM is Llama-3-8B-Instruct, and results are averaged over five runs.

Baselines

  • Oversampling and SMOTE-style methods, GraphSMOTE, MixupForGraph, LTE4G, HierTail, and three SAVE-TAG variants.

Reported results

  • BEAST is best in 22 of 24 dataset–metric cells. It loses Computer bAcc and Children F1 (Tables 1–2).
  • An ablation removes each component in turn (Fig. 3).
  • Generation strategies are compared before structural modeling (Table 3).
  • A 2-D visualization compares SMOTE and BEAST samples on Cora (Fig. 4a).
  • Sensitivity to τ_bd and τ_e is shown (Fig. 5).
  • An edge-recovery case study covers six low-degree nodes (Fig. 4b).
Critical strengths

[S1] The design is coherent and controllable.

  • Each of the two stated failure modes (Fig. 1) maps to a component.
  • Generation is conditioned explicitly on the competing classes (Eqs. 2–5), not on the tail class alone.
  • The acceptance rules (the Eq. 4 window, the Eq. 7 margin and the LLM critic) make the "boundary-aware" intent concrete and tunable. This is a sensible step beyond unconstrained LLM generation or SMOTE-style interpolation.
  • Fig. 2 makes the pipeline easy to follow.

[S2] Topology for synthetic nodes is treated as a first-class problem.

  • Node-synthesis methods often handle this weakly, either with a jointly trained edge generator (GraphSMOTE) or with confidence-based attachment (SAVE-TAG).
  • BEAST learns a scorer from structural responses and combines it with semantic similarity (Eq. 11). It also enforces a minimum degree that removes unsupported synthetic nodes (Eq. 15).
  • The "w/o Unbiased Topology Imputation" ablation (Fig. 3; Cora Acc 77.6 → ≈ 76) and Fig. 4b give initial evidence that this matters.
Critical weaknesses

[W1] The central tail-class claim is not directly evidenced, and most margins on the imbalance-sensitive metrics are within noise.

  • Issue: The point here is not the size of the SOTA margin. It is whether the evidence supports the stated mechanism, better tail-class recognition (Obs. ❷).
    • Against the best baseline in each column, BEAST's margin exceeds the larger of the two standard deviations in only 6 of 24 cells: Acc on Cora, Citeseer, Photo and Computer, and F1 on Citeseer and Photo.
    • On GMean and bAcc, the metrics Obs. ❷ relies on, none of the 12 margins exceeds the larger standard deviation. On Citeseer (GMean and bAcc) and Children (bAcc), the margin exactly equals it. Examples: PubMed bAcc 77.8 ± 0.6 vs. 77.6 ± 0.7; Photo GMean 84.6 ± 0.6 vs. 84.4 ± 0.2; and Computer bAcc, where BEAST loses (72.4 ± 2.4 vs. 72.7 ± 0.6).
    • On Photo, Accuracy rises by 10.5 points but bAcc by only 0.6. Acc weights per-class recall by test frequency and bAcc does not, so this suggests the gain is concentrated in classes that are frequent in the test set. Whether those are head or tail classes cannot be determined without a per-class breakdown.
    • Only five runs are reported, and it is not stated whether the LLM generation is re-sampled for each run. If it is generated once, the reported standard deviations understate the true variance.
  • :Why it matters: Aggregate metrics with overlapping error bars cannot establish better tail-class learning, and the Photo pattern is at least compatible with a different mechanism.

[W2] The gains may come from the loss and inference rule rather than from generation or topology, and the simplest strong baselines are missing.

  • Issue
    • Eq. 17 adds log ν, the log label frequency, to the logits. This is logit adjustment (Menon et al., "Long-tail learning via logit adjustment", ICLR 2021), which is closely related to Balanced Softmax (Ren et al., NeurIPS 2020).
    • It is a strong long-tail technique on its own. It is not cited, and the main text does not say whether any baseline uses it.
    • In Fig. 3, removing hierarchical supervision drops BEAST to ≈ 73.3 (Cora), ≈ 73.4 (PubMed) and ≈ 69.9 (Citeseer). That is below SAVE-TAG_S on Cora and PubMed (76.4 and 76.3) and level with it on Citeseer (69.9).
    • To be fair to the authors, the same holds when any single component is removed on Cora and PubMed (e.g., w/o topology ≈ 76.2 and ≈ 74.9). So the ablation shows that all components are needed. The log ν term, however, is never isolated.
    • Two further confounds:
      • The parent aggregation (Eq. 19) gives each tail class a log-sum-exp over 1+∣Act∣1 + |A_{c_t}| logits. This may act as a prior toward tail classes, although LparL_{par} (Eq. 20) trains through the same aggregation.
      • The generation budget is small. At IR = 0.2 the deficit is 16, so each tail class gets at most ⌈16γ⌉⌈16γ⌉ candidates before filtering (§3.3, §4.1). A +10.5 Acc gain on Photo from so few nodes deserves an explanation.
    • Several cited and directly relevant imbalance baselines are not compared:
      • GraphENS (Park et al.);
      • TAM (Song et al., ICML 2022);
      • GraphSHA (Li et al., KDD 2023);
      • ReNode (Chen et al., NeurIPS 2021);
      • BAT (Liu et al., ICML 2024; not cited).
    • Tables 1–2 also contain no plain GCN or MLP, with or without re-weighting or logit adjustment.
  • Why it matters: If much of the improvement comes from logit adjustment and the aggregation rule, the contribution of the LLM and topology components, which are the paper's main novelty, is overstated.
Questions for authors
  1. For each dataset, what are ∣Vsh∣|V_{sh}|, ∣Esh∣|E_{sh}| and ∣Du∣|D_u|?
    • Does the shadow graph contain unlabeled nodes? If so, where do their labels yiy_i come from, in Eq. 8 and in the cross-entropy loss of Eq. 9 (pseudo-labels, or validation/test labels)?
    • How are A^−Â⁻ and the χχ features defined on GprovG_{prov}?
    • How does the scorer behave on generated–original pairs when E_init is removed before scoring?
  2. How were hyperparameters chosen: on which split, with what grid, and with what budget for the baselines? Is Fig. 5 computed on the test set?
  3. What exactly does "w/o Hierarchical Supervision" remove? Please report:
    • (a) GCN + logit adjustment (or Balanced Softmax);
    • (b) SAVE-TAG_S + the §3.5 loss;
    • (c) BEAST with log ν but flat labels (no subclasses). $. Please give head-class vs. tail-class recall/F1 for Tables 1–2. Which classes drive the Photo result (+10.5 Acc vs. +0.6 bAcc)? Is the LLM generation re-sampled for each of the five runs?
  4. How many nodes are generated, and how many are accepted at each filtering stage (Eq. 4, Eq. 7, the critic, K_min)?
    • What fraction of accepted texts has cosine similarity > 0.95, or substantial n-gram overlap, with any test-node text?
    • How do results change when those texts are removed?
  5. On all masked tail edges, what are the precision, recall and AUC of the edge scorer, compared with the "semantic kNN" method of Fig. 4b and a VGAE or SEAL predictor? How is "semantic kNN" defined, and is it the same rule as E_init in Eq. 13? What are the head-class neighbour ratio and the homophily of E⁺?
4 comments
AuthorsOpenReview ↗

Response 1

We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following.

W1: Tail-class Evidence, Photo Attribution and Generation Variance. R1: Thank you for your comment. We address the your concerns from the following aspects: Photo Attribution. The per-class results demonstrate BEAST's ability to recognize a diverse range of tail categories, while maintaining strong performance on most classes. The results also reveal variations across individual categories, suggesting potential room for further improvement.

ClassGroupBEAST (%)
Video SurveillanceTail93.54 ± 0.75
AccessoriesTail49.79 ± 2.49
Binoculars & ScopesTail94.62 ± 1.53
VideoTail61.61 ± 3.20
Lighting & StudioTail73.58 ± 1.71
Bags & CasesTail75.73 ± 2.43
Tripods & MonopodsTail82.27 ± 1.53
FlashesTail65.08 ± 1.73
Digital CamerasHead86.83 ± 1.78
Film PhotographyHead83.83 ± 3.48
LensesHead82.79 ± 1.89
Underwater PhotographyHead28.37 ± 5.28

Generation Variance. LLM-generated data are independently regenerated in each run, so the reported standard deviations reflect variability from both data generation and model training.

Tail-class Evidence. To further assess BEAST's effectiveness across class frequencies, we evaluate head- and tail-class accuracy separately. As shown in the table, BEAST improves tail accuracy by 4.62 and 8.07 points on PubMed and Citeseer, respectively. Notably, Citeseer exhibits substantial tail-class gains with only marginal head-class improvement. While both improve on PubMed, tail accuracy decreases slightly by 0.61 points on Cora. These results provide more direct evidence of BEAST's benefits for tail-class recognition, although the effect varies across datasets.

DatasetMethodHead Acc (%)Tail Acc (%)
CoraSAVE-TAG-S82.7175.51
BEAST84.7274.90
PubMedSAVE-TAG-S81.3859.71
BEAST84.6664.33
CiteseerSAVE-TAG-S61.8761.29
BEAST62.2969.36

W2: Logit adjustment, hierarchical loss, and missing baselines. R2: Thank you for your comment. We address the your concerns from the following aspects:

Logit adjustment and hierarchical aggregation. BEAST employs a hierarchical learning objective that jointly optimizes leaf-level discrimination and parent-level classification. Specifically, Eq. (17) combines a label-frequency prior (log⁡ν\log\nu) with label-dependent silencing (log⁡λsMyi\log\lambda_s M_{y_i}) to alleviate class imbalance and parent–child competition. Eq. (19) then aggregates auxiliary subclass logits into their corresponding parent classes via LogSumExp, while Eq. (20) provides complementary parent-level supervision. At inference, predictions are made in the original label space using the aggregated parent logits. To quantify the contribution of frequency adjustment, we remove only log⁡ν\log\nu while keeping other components unchanged, resulting in Macro-F1 drops of 1.01, 1.24, and 0.97 percentage points on Cora, PubMed, and Citeseer, respectively.

Additional baselines. We added GCN, GCN+LA, GraphENS[1], TAM[2], GraphSHA[3], ReNode[4], and BAT[5] across all six datasets, using same backbone, data split, and checkpoint selection criterion. Results are averaged over five seeds. Overall, BEAST achieves the best or tied-best Macro-F1 on five of the six datasets, demonstrating its effectiveness across diverse long-tailed graph benchmarks.

MethodCoraPubMedCiteseerPhotoComputerChildren
GCN71.98 ± 0.4550.37 ± 6.0047.08 ± 0.5857.40 ± 1.0146.37 ± 0.773.42 ± 0.91
GCN + LA74.24 ± 0.4571.83 ± 0.3160.69 ± 1.0862.27 ± 0.5553.88 ± 0.5417.07 ± 0.87
TAM68.27 ± 0.7459.62 ± 3.8150.88 ± 0.9754.93 ± 1.0841.98 ± 1.633.42 ± 0.91
GraphENS74.05 ± 0.6371.90 ± 0.5361.23 ± 1.7664.13 ± 1.1260.34 ± 0.9619.89 ± 0.24
GraphSHA75.30 ± 1.1272.10 ± 1.0856.03 ± 1.5860.59 ± 0.4856.01 ± 0.8611.78 ± 0.47
ReNode74.58 ± 0.6669.64 ± 1.5059.06 ± 1.6059.85 ± 0.4948.86 ± 1.1612.45 ± 1.61
BAT74.10 ± 0.3370.76 ± 0.9455.73 ± 5.6558.95 ± 0.8649.63 ± 1.223.72 ± 1.00
BEAST75.3 ± 0.376.7 ± 0.168.9 ± 0.764.6 ± 0.553.3 ± 0.623.8 ± 1.2

Generation and topology. At IR = 0.2, Photo has eight tail classes with 16 missing labels each. We generated 128 candidate texts and retained 97 nodes after filtering. Boundary-oriented sampling achieves 74.20 Macro-F1 on Cora, compared with 74.12 for random sampling and 74.00 for interior sampling. These results suggest complementary contributions from semantic generation and topology recovery, beyond the gains from logit adjustment.

AuthorsOpenReview ↗

Response 2

Q1: Shadow Graph Construction, Provisional scoring R3: Thank you for your comment. We address the your concerns from the following aspects: Shadow Graph Construction. We adopt a transductive setting, where the full graph topology and node features are accessible, but ground-truth labels from validation and test nodes are never used. We first train an initial GNN using only labeled training nodes for supervision, and then generate pseudo-labels for unlabeled nodes. These pseudo-labels, together with ground-truth training labels, are used for class-aware sampling in Eq. (8) and cross-entropy training in Eq. (9). Thus, the shadow graph may contain unlabeled validation/test nodes, but no validation/test labels are involved.

DatasetNumber of Nodes (∣Vsh∣)(\lvert V_{\mathrm{sh}}\rvert)Number of Edges (∣Esh∣)(\lvert E_{\mathrm{sh}}\rvert)
Cora1,6241,817
PubMed11,83017,773
Citeseer1,9111,540

In the shadow graph, A^−\hat A^{-} denotes the normalized adjacency matrix after removing the sampled edges DuD_u. The confidence-response feature χi\chi_i captures the two-hop changes in node confidence through adjacency propagation, as defined in Eq. (10). For Gprov=(V∪V+,E∪Einit)G_{\mathrm{prov}}=(V\cup V^+, E\cup E_{\mathrm{init}}), we apply the frozen shadow model to obtain provisional posteriors ρi,prov\rho_{i,\mathrm{prov}} and confidence trends χi,prov\chi_{i,\mathrm{prov}}, which are then used to construct the pairwise features in Eq. (11).

Provisional scoring. To examine the scorer's behavior when EinitE_{\mathrm{init}} is removed before scoring generated–original pairs, we conduct an ablation experiment. As shown in the table, removing EinitE_{\mathrm{init}} reduces performance by 0.5, 2.1, and 4.1 percentage points on Cora, PubMed, and Citeseer, respectively. This indicates that the scorer can still evaluate generated–original pairs using the remaining structural information, but retaining EinitE_{\mathrm{init}} provides valuable structural context and leads to more effective pair selection.

MethodCoraPubMedCiteseer
Remove EinitE_{\mathrm{init}}74.8 ± 0.274.6 ± 0.564.8 ± 1.1
BEAST75.3 ± 0.376.7 ± 0.168.9 ± 0.7

Q2: Hyperparameter Selection R4: Thank you for your comment. All hyperparameters were selected based on the validation set, while the test set was reserved exclusively for final evaluation. For BEAST, we examined τbd\tau_{bd} and τe\tau_e, as shown in Fig. 5. For all baselines, we used the same data splits and validation-based model selection protocol, with their respective hyperparameter search spaces and tuning budgets explicitly reported. Figure 5 is computed on the validation set, not the test set. We will clarify these details in the revised manuscript.

Q3: Hierarchical supervision, subgroup results, and generation seeds

R5: Thank you for your comment. We address the your concerns from the following aspects:

Clarification of "w/o Hierarchical Supervision." This variant retains BEAST's generated nodes and recovered topology but removes the hierarchical learning mechanism in Section 3.5. Specifically, auxiliary subclass labels are mapped to their original parent classes, and flat classification is performed over the original label space. We retain frequency-based logit adjustment (log⁡ν\log\nu) while removing subclass-specific supervision, parent–child silencing, and parent-level aggregation and CE.

Controlled comparisons We provide three additional comparisons. As shown in the table, the full BEAST achieves the best Macro-F1 across all three datasets (75.30%, 76.70%, and 68.90%), validating the effectiveness of integrating fine-grained semantic supervision with hierarchical learning.

Variant (Macro-F1, %)CoraPubMedCiteseer
GCN + LA, original graph74.4071.8260.58
SAVE-TAG-S + flat CE73.4172.2443.99
SAVE-TAG-S + hierarchical loss72.6973.0447.91
BEAST, no log-frequency adjustment74.1975.4667.93
BEAST, flat labels + LA68.8469.4959.49
BEAST, flat labels + CE70.0665.0758.49
BEAST, no parent CE74.1972.2865.02
BEAST, no silencing74.9671.9265.73
BEAST, randomized subclass assignments74.3072.0865.09
BEAST75.376.768.9

Head and tail-class performance. Please refer to our response to W1 for the detailed head-class and tail-class Recall/F1 results.

Per-class Accuracy on Photo Please refer to our response to W1.

LLM generation across runs. LLM-generated texts are independently regenerated for each of the five random seeds; they are not shared across runs.

AuthorsOpenReview ↗

Response 3

Q4: Generation, Filtering, and Test-text Overlap R6: Thank you for your comment. We address the your concerns from the following aspects: Generation and filtering statistics. We report the archived pipeline statistics and distinguish subclass proposals from generated nodes. Eq. (4) filters auxiliary subclass proposals, while Eq. (7) and the LLM critic jointly determine semantic acceptance. The archived logs preserve the joint strict-pass counts but do not separately record the rejection counts of Eq. (7) and the critic. The subsequent topology filtering retains 79/80, 30/30, and 37/37 nodes on Cora, PubMed, and Citeseer, respectively, as detailed in the table. We also report a separate fixed-budget connection control, which uses four undirected edges per accepted node and does not apply the original Kmin⁡K_{\min} filtering rule. Since no generated texts were flagged, removing the potentially overlapping texts leaves the augmented graphs unchanged.

DatasetRetained subclass proposalsCached generated textsAccepted for the fixed-budget control
Cora1014064
PubMed35621
Citeseer611241

Test-text overlap. We audited accepted generated texts against all test-node texts using two criteria: maximum sentence-embedding cosine similarity (>0.95) or maximum trigram Jaccard overlap (≥\geq 0.50). No accepted texts were flagged (0/64 on Cora, 0/21 on PubMed, and 0/41 on Citeseer). This audit was conducted post hoc and was not used for model training or selection.

DatasetAccepted Generated TextsCosine Similarity > 0.95Trigram Jaccard Similarity ≥ 0.50Flagged by Either Criterion
Cora64000/64 (0%)
PubMed21000/21 (0%)
Citeseer41000/41 (0%)
Total126000/126 (0%)

Q5: Masked-edge Recovery and Added-edge Composition R7: Thank you for your comment. We address the your concerns from the following aspects: Edge Recovery Performance. We randomly masked 20% of eligible tail-incident undirected edges and evaluated three independent masks, with 837, 5,952, and 740 positive pairs per mask on Cora, PubMed, and Citeseer, respectively. Sampled nonedges were used at a 1:10 ratio. The retrieval budget equaled the number of positives, making Precision and Recall identical.

DatasetMethodPrecision = RecallROC-AUCAverage Precision
CoraBEAST0.64360.94550.7079
Semantic kNN0.62840.92640.6921
VGAE0.50900.88860.4770
PubMedBEAST0.76300.97110.8299
Semantic kNN0.70530.95290.7615
VGAE0.72180.95950.7782
CiteseerBEAST0.73650.97000.8083
Semantic kNN0.71980.96290.7916
VGAE0.54460.90850.5322

Semantic kNN Definition. Semantic kNN shares the similarity principle of Eq. (13), but serves as an independent retrieval baseline rather than an initialization step. Head-class Neighbor Ratio and Homophily The statistics are computed after graph construction using original endpoint labels. Head-class neighbor ratio measures the proportion of original neighbors belonging to head classes, while same-parent homophily measures agreement between generated nodes' assigned parent labels and their original neighbors' labels.

DatasetAdded EdgesHead-class Neighbor RatioSame-parent Homophily
Cora25618.36%59.38%
PubMed842.38%90.48%
Citeseer16414.02%53.05%

Reference

[1] Joonhyung Park, et al.: GraphENS: Neighbor-Aware Ego Network Synthesis for Class-Imbalanced Node Classification. ICLR 2022.

[2] Jaeyun Song, et al.: TAM: Topology-Aware Margin Loss for Class-Imbalanced Node Classification. ICML 2022.

[3] Wen-Zhi Li, et al.: GraphSHA: Synthesizing Harder Samples for Class-Imbalanced Node Classification. KDD 2023.

[4] Deli Chen, et al.: Topology-Imbalance Learning for Semi-Supervised Node Classification. NeurIPS 2021.

[5] Zhining Liu, et al.: Class-Imbalanced Graph Learning without Class Rebalancing. ICML 2024.

In light of these responses, we hope we have addressed your concerns, and hope you will consider raising your score. If there are any additional notable points of concern that we have not yet addressed, please do not hesitate to share them, and we will promptly attend to those points.

Dear Reviewer Duxe,

We thank you for your careful reviews and constructive suggestions. With the rebuttal-discussion period coming to a close, may we kindly ask you to review them and confirm if they have effectively addressed your concerns? Any further insights or comments you wish to share would be warmly welcomed. We extend our heartfelt gratitude for the time and effort you've invested in reviewing our work.

Best regards,

The Authors

Reviewer_BswEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses long-tailed node classification on text-attributed graphs by combining tail-class text generation with graph structure completion. The approach is well motivated, addressing both sample scarcity and structural bias, and shows notable gains on some datasets. However, the individual contributions of the new components have not been sufficiently validated, and inconsistencies in some reported metrics require clarification.

Critical strengths
  1. Node classification on graphs relies on neighborhood information propagation, so both the semantic content and connectivity of generated samples affect classification performance. The authors therefore combine tail-class semantic augmentation with topology completion, predicting connections for new nodes as they add tail-class samples. The method also uses parent-class supervision to keep auxiliary-class learning aligned with the original classification objective. Together, these components form a coherent design.
  2. The paper evaluates the method on six datasets, with particularly notable results on Photo. It also examines different imbalance levels, GNN backbones, and LLM configurations, providing further insight into how model choice and the degree of class imbalance affect performance.
Critical weaknesses
  1. Table 2 reports a GMean of 51.7+-2.4 and a bAcc of 35. 0+-1.1 for MixupForGraph on Children. However, these statistics are incompatible with the constraints implied by Eqs. (33)–(35) in a 24-class single-label setting. The authors need to verify the metric implementation and reported values, as this inconsistency undermines confidence in the experimental comparisons.
  2. The edge predictor is trained to recover masked edges between real nodes but is then used to connect nodes generated by an LLM. Since the generated nodes may differ in their semantic distributions and initial neighborhoods, the existing analysis on real nodes does not sufficiently establish the reliability of this transfer. Direct validation of the connections assigned to generated nodes is therefore needed.
  3. Figure 3 does not clearly describe the implementation of each ablation or report the resulting numbers of retained generated nodes and added edges. Consequently, it is difficult to determine whether the performance changes arise from the proposed mechanisms or from changes in the amount of augmentation. The results also do not sufficiently isolate the contributions of boundary filtering and parent-class supervision.
Questions for authors
  1. Were the GMean and bAcc values for MixupForGraph on Children in Table 2 computed according to Eqs. (33)–(35)? How can the inconsistency between these values be explained?
  2. What direct evidence shows that the edge predictor trained on real nodes can reliably establish connections for new nodes generated by an LLM?
  3. How is each ablation in Figure 3 implemented, and what are the resulting numbers of retained generated nodes and added edges? Can boundary filtering be compared under matched augmentation budgets, and parent-class supervision evaluated using the same augmented graph?
  4. In Section 4.1, does “using only training nodes” mean that only training labels are used for supervision, or that the graph itself contains only training nodes? Please specify the actual numbers of nodes and edges in the shadow graph. If unlabeled nodes are included, how is the label-dependent sampling in Eq. (8) implemented?
3 comments
AuthorsOpenReview ↗

Response 1

We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following.

W1&Q1: MixupForGraph metric inconsistency on Children

W1: Table 2 reports a GMean of 51.7+-2.4 and a bAcc of 35. 0+-1.1 for MixupForGraph on Children. However, these statistics are incompatible with the constraints implied by Eqs. (33)–(35) in a 24-class single-label setting. The authors need to verify the metric implementation and reported values, as this inconsistency undermines confidence in the experimental comparisons.

Q1: Were the GMean and bAcc values for MixupForGraph on Children in Table 2 computed according to Eqs. (33)–(35)? How can the inconsistency between these values be explained?

R1: Thank you for your comment. Upon double-checking, we identified typos in the Children results of MixupForGraph and HierTail and corrected the corresponding entries in Table 2. In particular, the corrected GMean for MixupForGraph is 56.7%.

W2&Q2 Transfer of Edge Eecovery to Generated Nodes.

W2: The edge predictor is trained to recover masked edges between real nodes but is then used to connect nodes generated by an LLM. Since the generated nodes may differ in their semantic distributions and initial neighborhoods, the existing analysis on real nodes does not sufficiently establish the reliability of this transfer. Direct validation of the connections assigned to generated nodes is therefore needed.

Q2:What direct evidence shows that the edge predictor trained on real nodes can reliably establish connections for new nodes generated by an LLM?

R2: Thank you for your comment. To evaluate whether the edge predictor generalizes to LLM-generated nodes, we conduct two complementary experiments. Cold-start Edge Recovery. As shown in the Table, we hide all incident edges of 20% of test tail nodes and evaluate neighbor retrieval using three methods. Recovery achieves the highest AUC across all three datasets and consistently outperforms VGAE on all metrics. Although Semantic kNN performs slightly better on two metrics on Cora, Recovery achieves competitive overall performance, demonstrating its ability to generalize to unseen nodes. The shadow graph size is presented in the table below.

DatasetMethodAUC ↑AP ↑P=R ↑
CoraBEAST0.93490.68040.6291
Semantic kNN0.92660.69790.6340
VGAE0.83990.36720.4026
PubMedBEAST0.96130.78420.7280
Semantic kNN0.95530.77010.7088
VGAE0.93840.66460.6277
CiteseerBEAST0.96340.79510.7175
Semantic kNN0.96180.79380.7158
VGAE0.85480.39030.4317

Downstream Evaluation on Generated Nodes. As shown in the Table, we fix the synthetic texts, subclass labels, downstream loss, and data split, changing only the connection strategy. Recovery consistently outperforms Semantic kNN across all three datasets, improving Macro-F1 by 0.48, 0.67, and 1.50 percentage points on Cora, PubMed, and Citeseer, respectively. These results demonstrate the practical benefits of the proposed recovery strategy when connecting LLM-generated nodes.

DatasetSemantic kNNBEASTImprovement
Cora73.8274.30+0.48
PubMed71.5072.17+0.67
Citeseer63.6065.10+1.50
AuthorsOpenReview ↗

Response 2

W3&Q3: Ablation Details and Fine-grained Ablations

W3: Figure 3 does not clearly describe the implementation of each ablation or report the resulting numbers of retained generated nodes and added edges. Consequently, it is difficult to determine whether the performance changes arise from the proposed mechanisms or from changes in the amount of augmentation. The results also do not sufficiently isolate the contributions of boundary filtering and parent-class supervision.

Q3: How is each ablation in Figure 3 implemented, and what are the resulting numbers of retained generated nodes and added edges? Can boundary filtering be compared under matched augmentation budgets, and parent-class supervision evaluated using the same augmented graph?

R3: Thank you for your comment. In the original ablation study, we construct four variants by modifying the corresponding components of BEAST:

  1. w/o Auxiliary Subclass Proposal: We directly prompt the LLM to generate synthetic nodes for each tail class, without first discovering auxiliary subclasses to guide generation.
  2. w/o Semantic Extrapolation: We remove the subclass-guided semantic extrapolation process, so no boundary-oriented synthetic nodes are generated and incorporated into the graph.
  3. w/o Unbiased Topology Imputation: We retain the generated nodes and their semantic information but disable the proposed shadow-model-based edge recovery for connecting them to the original graph.
  4. w/o Hierarchical Supervision: We remove the auxiliary subclass-level supervision and train the model using only the original parent-class classification objective, without hierarchical label supervision.

We further compare boundary-oriented sampling with Random and Interior sampling while keeping the number of accepted synthetic nodes fixed. As shown in the Table, BEAST consistently achieves the best performance across all five metrics, demonstrating the effectiveness of boundary-aware filtering in selecting informative synthetic nodes.

MethodAccuracy (%)Macro-F1 (%)Balanced accuracy (%)Tail macro recall (%)Tail macro-F1 (%)
Random75.57 ± 0.2573.12 ± 0.1574.67 ± 0.1571.44 ± 0.3972.30 ± 0.33
Interior75.48 ± 0.2573.00 ± 0.3573.74 ± 0.1771.29 ± 0.4972.12 ± 0.53
BEAST76.63 ± 0.6074.20 ± 0.4575.99 ± 0.4872.77 ± 0.4272.44 ± 0.56

Shadow Graph Size and Label Access

Q4: In Section 4.1, does “using only training nodes” mean that only training labels are used for supervision, or that the graph itself contains only training nodes? Please specify the actual numbers of nodes and edges in the shadow graph. If unlabeled nodes are included, how is the label-dependent sampling in Eq. (8) implemented?

R4: Thank you for your comment. We adopt a transductive setting, where the full graph topology and node features are accessible, but ground-truth labels from validation and test nodes are never used. We first train an initial GNN using only labeled training nodes for supervision, and then generate pseudo-labels for unlabeled nodes. These pseudo-labels, together with ground-truth training labels, are used for class-aware sampling in Eq. (8) and cross-entropy training in Eq. (9). Thus, the shadow graph may contain unlabeled validation/test nodes, but no validation/test labels are involved.

DatasetNumber of Nodes (∣Vsh∣)(\lvert V_{\mathrm{sh}}\rvert)Number of Edges (∣Esh∣)(\lvert E_{\mathrm{sh}}\rvert)
Cora1,6241,817
PubMed11,83017,773
Citeseer1,9111,540

In light of these responses, we hope we have addressed your concerns, and hope you will consider raising your score. If there are any additional notable points of concern that we have not yet addressed, please do not hesitate to share them, and we will promptly attend to those points.

Dear Reviewer BswE,

We thank you for your careful reviews and constructive suggestions. With the rebuttal-discussion period coming to a close, may we kindly ask you to review them and confirm if they have effectively addressed your concerns? Any further insights or comments you wish to share would be warmly welcomed. We extend our heartfelt gratitude for the time and effort you've invested in reviewing our work.

Best regards,

the Authors

Reviewer_invjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BEAST for long-tailed text-attributed graph learning. The method uses LLMs to generate auxiliary subclasses and synthesize tail-class nodes near semantic decision boundaries. It then uses a class-rebalanced shadow graph to learn connectivity patterns and recover edges for the generated nodes. Finally, hierarchical supervision combines subclass-level and parent-class learning. Experiments on six text-attributed graph benchmarks show that BEAST generally outperforms existing graph augmentation and long-tailed learning methods.

Critical strengths
  1. The paper addresses an important problem in long-tailed graph learning. It considers both the limited semantic information of tail classes and the biased graph structure, providing a more complete view of class imbalance.
  2. The proposed method is well motivated. LLM-guided auxiliary subclasses and boundary-aware semantic extrapolation provide a reasonable way to generate more diverse and informative samples for tail classes.
  3. The topology imputation module is a useful part of the framework. Instead of relying only on semantic similarity, it learns connectivity patterns to reduce incorrect edges when adding generated nodes to the graph.
  4. The experiments are relatively comprehensive, covering six datasets and several representative baselines. The ablation studies, imbalance-ratio analysis, and experiments with different LLM and GNN backbones provide good support for the effectiveness of the method.
Critical weaknesses
  1. The framework includes several complex stages, such as LLM-based generation, semantic filtering, shadow-model training, and topology imputation. However, the paper does not provide a clear analysis of the additional computational cost and training time. It is difficult to judge the practical efficiency of the method.
  2. The quality of the generated tail-class samples is not independently evaluated. Although the authors use prototype-based filtering and an LLM critic, these steps may still introduce incorrect or unrealistic samples, especially near class boundaries.
  3. The reliability of topology imputation needs stronger validation. The current edge-recovery analysis mainly uses a small number of cases. In addition, the theoretical bound assumes that predicted edge scores are close to true edge probabilities, but this assumption is not directly verified in the experiments.
  4. The experiments mainly use artificially constructed long-tailed training sets. Although different imbalance ratios are tested, it remains unclear whether the proposed method performs equally well on naturally imbalanced graphs with more complex structural patterns.
Questions for authors
  1. Could the authors provide a comparison of the computational cost, training time, and overall efficiency of BEAST against SAVE-TAG and simpler augmentation methods?
  2. Could the authors provide a more comprehensive evaluation of the quality of generated tail-class samples and recovered graph edges, beyond the current qualitative analysis?
  3. How well does BEAST generalize to naturally imbalanced graphs, rather than datasets where class imbalance is created by reducing the number of labeled nodes?
3 comments
AuthorsOpenReview ↗

Response 1

We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following.

W1&Q1. Computational Cost and Runtime.

W1: The framework includes several complex stages, such as LLM-based generation, semantic filtering, shadow-model training, and topology imputation. However, the paper does not provide a clear analysis of the additional computational cost and training time. It is difficult to judge the practical efficiency of the method.

Q1: Could the authors provide a comparison of the computational cost, training time, and overall efficiency of BEAST against SAVE-TAG and simpler augmentation methods?

R1: Thank you for your comment. As shown in the table, we separately measure the preparation and classifier training times on the same machine, with all times reported in seconds. Training times are reported as averages over five random seeds. Although BEAST incurs additional preparation costs due to LLM-based generation and filtering, it achieves improved performance. These costs can be reduced through caching, while classifier training remains comparable to existing baselines.

DatasetBEAST PreparationSAVE-TAG-S PreparationGCN + LAWeighted oversamplingSAVE-TAG-SBEAST
Cora575.46147.101.431.311.231.43
PubMed277.4898.471.421.431.351.58
Citeseer722.97149.761.231.231.211.36

We also analyze the computational complexity of BEAST and SAVE-TAG-S to complement the empirical runtime comparison. Their overall complexities are summarized below: BEAST

O\bigl(&G_B+X_B+d[N_{\mathrm{tr}}+C^2+(U+R)(1+q)+N] +N\log N+M\\ &+(T_{\mathrm{sh}}+r)F(N_{\mathrm{sh}},M_{\mathrm{sh}}) +Q(d+C+T_e g)\\ &+S_BNd+F(N+S_B,M+S_BK_{\mathrm{init}}) +P(d+C+g)\\ &+T_{\mathrm{cls}}F(N_{\mathrm{aug}},M_{\mathrm{aug}})\bigr). \end{aligned}$$ **SAVE-TAG-S** $$\begin{aligned} O\bigl(&G_S+X_S+d\textstyle\sum_c n_c^2 +T_{\mathrm{cf}}Q_{\mathrm{cf}}(d+1)\\ &+d(N+S_S)+S_S(N+S_S) +T_{\mathrm{cls}}F(N_{\mathrm{aug}},M_{\mathrm{aug}})\bigr). \end{aligned}$$ Although BEAST introduces additional shadow graph learning costs, it shares the same classifier training complexity as SAVE-TAG-S while avoiding exhaustive pairwise scoring through staged candidate filtering. We further compare the LLM usage of BEAST and SAVE-TAG in terms of invocation counts and token consumption. Although BEAST requires more LLM calls and incurs comparable token consumption, it achieves better classification performance, demonstrating the effectiveness of its more comprehensive semantic augmentation. | Dataset | BEAST LLM calls | BEAST input / output tokens | SAVE LLM calls | SAVE input / output tokens | |---|---:|---:|---:|---:| | Cora | 32 | 39,664 / 19,897 | 20 | 36,331 / 16,711 | | PubMed | 15 | 27,639 / 10,355 | 8 | 29,041 / 9,541 | | Citeseer | 21 | 43,129 / 23,838 | 16 | 33,476 / 17,300 | Overall, BEAST introduces additional one-time computational overhead due to its LLM-based semantic augmentation, while maintaining comparable training time. **W2&Q2. Quality of Generated Tail-class Nodes.** > W2: The quality of the generated tail-class samples is not independently evaluated. Although the authors use prototype-based filtering and an LLM critic, these steps may still introduce incorrect or unrealistic samples, especially near class boundaries. > Q2: Could the authors provide a more comprehensive evaluation of the quality of generated tail-class samples and recovered graph edges, beyond the current qualitative analysis? R2: Thank you for your comment. We evaluate the semantic diversity and non-redundancy of the accepted generated texts using cosine similarity and trigram Jaccard similarity. Among the 126 accepted samples, none exhibits a cosine similarity above 0.95 or a trigram Jaccard similarity of at least 0.50 with existing test-node texts. These results indicate that the generated samples are not merely superficial variations or repetitions of existing texts, providing additional non-redundant semantic content for tail classes. | Dataset | Accepted Generated Texts | Cosine Similarity > 0.95 | Trigram Jaccard Similarity ≥ 0.50 | Flagged by Either Criterion | |---|---:|---:|---:|---:| | Cora | 64 | 0 | 0 | 0/64 (0%) | | PubMed | 21 | 0 | 0 | 0/21 (0%) | | Citeseer | 41 | 0 | 0 | 0/41 (0%) | | Total | 126 | 0 | 0 | 0/126 (0%) |
AuthorsOpenReview ↗

Response 2

W3&Q2. Quality and Calibration of Recovered Edges.

W2: The reliability of topology imputation needs stronger validation. The current edge-recovery analysis mainly uses a small number of cases. In addition, the theoretical bound assumes that predicted edge scores are close to true edge probabilities, but this assumption is not directly verified in the experiments.

Q2: Could the authors provide a more comprehensive evaluation of the quality of generated tail-class samples and recovered graph edges, beyond the current qualitative analysis?

R3: Thank you for your comment. To further validate the reliability of topology imputation, we evaluate the calibration quality of recovered edge scores using Brier Score and Expected Calibration Error (ECE), comparing BEAST with Semantic kNN and VGAE. As shown in the Table, BEAST consistently achieves the lowest Brier Score and ECE across all three datasets. In particular, its ECE is substantially lower than both alternatives, indicating better alignment between predicted edge confidence and empirical edge frequencies. These results provide quantitative evidence that BEAST produces more reliable edge probability estimates, lending empirical support to the assumption underlying our theoretical error bound.

DatasetMetricBEASTSemantic kNNVGAE
CoraBrier0.04460.04670.0637
ECE0.00630.01450.0310
PubMedBrier0.03160.03870.0374
ECE0.00390.00530.0131
CiteseerBrier0.03430.03690.0591
ECE0.00750.01640.0308

W4&Q3. Generalization under Natural Class Imbalance.

W4: The experiments mainly use artificially constructed long-tailed training sets. Although different imbalance ratios are tested, it remains unclear whether the proposed method performs equally well on naturally imbalanced graphs with more complex structural patterns.

Q3: How well does BEAST generalize to naturally imbalanced graphs, rather than datasets where class imbalance is created by reducing the number of labeled nodes?

R4: Thank you for your comment. To further evaluate BEAST under natural class imbalance, we conduct experiments on the Books-History dataset, an e-commerce text-attributed graph derived from Amazon-Books[1]. As shown in the Table, BEAST achieves comparable overall accuracy to SAVE-TAG-S (80.73% vs. 80.61%), while substantially improving Macro-F1 and balanced accuracy by 5.24 and 13.31 percentage points, respectively. More notably, BEAST increases Tail Macro Recall from 1.51% to 20.93%, demonstrating a substantial improvement in recognizing underrepresented classes. These results highlight BEAST's effectiveness in alleviating the bias toward head classes and improving classification performance under naturally imbalanced class distributions.

MethodAccuracy (%)Macro-F1 (%)Balanced Accuracy (%)Tail Macro Recall (%)
SAVE-TAG-S80.61 ± 0.0235.61 ± 0.5134.04 ± 0.271.51 ± 0.59
BEAST80.73 ± 0.0440.85 ± 0.2147.35 ± 0.2820.93 ± 0.47

Reference [1] Hao Yan, et al.: A Comprehensive Study on Text-attributed Graphs: Benchmarking and Rethinking. NeurIPS 2023.

In light of these responses, we hope we have addressed your concerns, and hope you will consider raising your score. If there are any additional notable points of concern that we have not yet addressed, please do not hesitate to share them, and we will promptly attend to those points.

Dear Reviewer invj,

We thank you for your careful reviews and constructive suggestions. With the rebuttal-discussion period coming to a close, may we kindly ask you to review them and confirm if they have effectively addressed your concerns? Any further insights or comments you wish to share would be warmly welcomed. We extend our heartfelt gratitude for the time and effort you've invested in reviewing our work.

Best regards,

The Authors

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your thoughtful reviews and contributions. The authors have now submitted their rebuttals. Please take some time to review their responses and assess whether your concerns have been adequately addressed. If appropriate, please update your comments and ratings accordingly.

Thank you for your time and continued efforts.

Best regards,

Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.