acceptodds
Under review as a conference paper at ICLR 2027

Sample, Don't Search: One-Run Ensembles for Node Property Prediction

Abstract

Node property prediction often requires tuning graph propagation and model choices for each dataset. We introduce GraphPack, which explores these choices through graph-view sampling and ensemble selection in a single training run. A shared bank of propagated features and training-label views supplies different inputs to randomly configured MLPs. These MLPs train together as one packed module on a single GPU. Validation-based selection combines their snapshots with gradient-boosted trees, adapting the model and view mix to each graph under fixed sampling rules. On ten GraphLand datasets, GraphPack achieves the best or tied performance on seven against tree-based, tuned decoupled, and tuned message-passing pipelines with access to the same inputs. It also achieves higher mean scores on all ten datasets than an ensemble of the best configuration from a 50-trial SIGN search. That search-and-ensemble workflow takes 9.1 times the wall-clock time of GraphPack, measured by the geometric mean across datasets. Controlled ablations across six datasets show that distributing graph views across members improves over giving every member all views. These results identify graph-view diversity as a source of ensemble gains and support sampling and selection as a practical alternative to per-dataset tuning.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph neural networks, node classification, ensemble learning, hyperparameter optimization, graph foundation models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_9dwROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes GraphPack, a node-level prediction version of the recent method TabPack. It precomputes a shared bank of views (propagated features and training labels), samples view subsets and hyperparameters for 64 MLPs, and trains them efficiently as a packed module. Online greedy selection then combines MLPs and LightGBM candidates to improve the validation score and produce the final ensemble model. The goal of this method is to achieve performance similar to or better than tuned graph learning baselines without per-dataset hyperparameter tuning (search), which often takes substantial time. GraphPack achieved the best or tied performance on seven of ten GraphLand datasets under the RH split (random splits with 50% training labels), compared with the evaluated GNN, LightGBM, and SIGN pipelines. It also had a substantially lower measured workflow cost than tuning and ensembling the SIGN baseline.

Critical strengths
  • Ensembling is an effective, efficient, and widely used approach in tabular learning, but I believe its practical value is often overlooked in graph learning. Showing its effectiveness and efficiency in graph learning is a practical contribution of this work.
  • Ablations show that sampling view subsets brings benefits over giving every member the entire bank, supporting structural diversity as a contributor to the method’s performance. This highlights the importance of view diversity in building effective graph ensembles.
Critical weaknesses
  • Questionable transferability: This work uses 16 datasets for development, while only one web graph with three different tasks is used for held-out evaluation, if I understood correctly. While achieving strong performance across diverse datasets with a single method is difficult, I believe this experimental setup provides limited support for its generalizability and transferability to unseen graphs.
  • I believe the purpose of this method is to eliminate per-dataset hyperparameter tuning by ensembling inexpensive and diverse decoupled predictors. While the results are promising, graph foundation models, as discussed by the authors in the related work section, share a same target problem. Although the paper includes a comparison in Appendix D.4, I believe the cost and performance comparison should be discussed more prominently in the main experimental section.
Questions for authors
  • GraphPack performs worse than GNNs on the city-road datasets. Does this observation reflect fundamental limitations of decoupled GNNs? If GraphPack inherits these limitations, is its design also fundamentally limited on graphs where decoupled methods underperform?
  • GraphPack underperforms fine-tuned GraphPFN in the low-label regime. Do the authors have comparisons with GNNs on these same datasets and splits? If GraphPack also underperforms GNNs in the low-label regime, I believe this limitation should be adequately discussed in the main paper.
  • How did the authors ensure a fair comparison on the datasets used for development, which make up most of the evaluation in the paper? Doesn’t using these datasets for development give GraphPack an unfair advantage over the other baselines?
  • GNN baselines use the validation set for hyperparameter search and checkpoint selection, while GraphPack uses it for early stopping and online greedy ensemble selection. While it is hard to say which method benefits more from the validation set, I believe the difference in how it is used could affect the comparison. Can GraphPack still outperform the baselines when ensemble selection uses an inner split of the training set under a matched training-label budget? If so, this would provide stronger evidence.
5 comments
AuthorsOpenReview ↗

Response to Reviewer 9dwR (1/4)

Thank you for recognizing the practical value of ensembling in graph learning and the role of structural-view diversity. We have addressed each weakness and question with new experiments and focused manuscript revisions.

Three additions directly strengthen the contribution:

  • Transfer: fixed defaults work on two new social and financial graphs, extending held-out evaluation to three underlying graphs.
  • Low-label learning: GraphPack exceeds both locally tuned GAT variants and four published conventional GNN baselines on the two evaluated RL splits.
  • Selection budget: all validation-based adaptation fits inside the original training-label budget, with gains over plain GAT and a full-bank single MLP.

For easy navigation, W1, W2, and Q1–Q4 are answered separately below. New performance results are mean ± sample SD over three seeds. AP, ROC-AUC, R², and Spearman are multiplied by 100; BlueSky MAE uses log1p engagement units. Published results retain their source protocols.

W1 — Transfer to new graph domains

The fixed GraphPack procedure now covers three held-out graphs from different domains: web, social, and financial. The added BlueSky-quotes and Elliptic experiments test the same model defaults on previously unused graphs and improve on both newly trained controls.

For both added graphs, the model defaults and seeds were fixed before observing their effectiveness results, and all declared targets and seeds were retained. What transfers is the training procedure and its configuration distribution: predictors are fitted on the new graph, without transferring fitted predictor weights or conducting a new GraphPack hyperparameter search.

BlueSky-quotes: a new temporal social graph. We use the released GraphBench 0.2.0 features and chronological snapshots to predict log1p likes, replies, and reposts. The chronological train/validation/test partitions contain 289,136/286,425/311,272 node-snapshot observations. Propagation stays within each snapshot; label views use training labels only. All three targets and seeds are included.

The table summarizes aggregate MAE for our three locally evaluated methods and six published reference methods from GraphBench v5, Table 2, with the source identified for each row.

MethodSourceAggregate MAE ↓
GraphPackThis work0.7387 ± 0.0023
Fixed full-bank single MLPThis work0.8008 ± 0.0021
Raw-feature LightGBMThis work0.7882 ± 0.0001
DeepSetsGraphBench v50.810 ± 0.005
MLPGraphBench v50.784 ± 0.001
GNNGraphBench v50.768 ± 0.002
GIN+GraphBench v50.737 ± 0.014
GatedGCN+GraphBench v50.734 ± 0.010
GCN+GraphBench v50.781 ± 0.021

Our aggregate MAE averages the three targets within each seed, then reports the mean and sample SD across seeds. GraphPack improves on both local controls in every target-wise R², MAE, and Spearman score. Its aggregate MAE of 0.7387 is also below the published means of DeepSets, MLP, GNN, and GCN+, supporting the competitiveness of the fixed defaults on this previously unused temporal graph.

For likes, replies, and reposts, respectively, the 100R2100R^2 results are:

  • GraphPack: 23.77 ± 0.31 / 20.54 ± 0.37 / 23.94 ± 0.45.
  • Fixed full-bank single MLP: 15.94 ± 1.82 / 12.29 ± 1.52 / 16.57 ± 0.66.
  • Raw-feature LightGBM: 13.94 ± 0.02 / 10.82 ± 0.05 / 13.68 ± 0.03.

Elliptic: a new financial-transaction graph. We use chronological fit/validation/test periods 1–29/30–34/35–49 on Elliptic, with the same fixed model defaults and seeds 0/1/2.

The table also includes seven independently published references from GADBench (NeurIPS 2023), Tables 4 and 13. Its tuned-method values are point estimates under the source protocol; AUPRC is computed as average precision.

MethodSourceAP ↑ROC-AUC ↑
GraphPackThis work79.32 ± 0.2992.49 ± 0.89
Fixed full-bank single MLPThis work73.02 ± 3.3790.02 ± 0.84
LightGBM on released featuresThis work78.92 ± 0.0092.40 ± 0.00
GCNGADBench25.3781.69
GraphSAGEGADBench57.8287.59
BGNNGADBench62.0387.69
XGBoostGADBench76.9389.91
XGB-GraphGADBench77.7891.78
Random ForestGADBench78.4291.68
RF-GraphGADBench78.8693.21
AuthorsOpenReview ↗

Response to Reviewer 9dwR (2/4)

W1 — Transfer to new graph domains (continued)

GraphPack achieves 79.32 AP without an additional per-dataset search, exceeding both local controls and the seven published AP references shown here. Its mean ROC-AUC also exceeds both local controls. Elliptic's released features include neighborhood aggregates.

Why the fixed procedure can adapt to new graphs. The reusable rules specify which predictors to sample; training and selection determine how those predictors are combined for each graph. Figure 4 shows this adaptation under shared sampling rules: the selected view mix varies by graph, and selected label-view use tracks its measured benefit. The new social and financial results test this mechanism's practical value beyond the development tasks.

Manuscript revision: Sections 4.1 and 4.4 introduce the expanded transfer evaluation. Appendix H.1 (Tables 25–27) gives both new-graph protocols and results, including every BlueSky target and reported metric. Q3 below explains how this complements the development-benchmark comparisons.

W2 — Graph foundation models and cost in the main paper

The main paper now compares GraphPack with graph foundation models in accuracy and computational cost. Section 4.4 highlights the accuracy/cost comparison; Table 28 in Appendix H.2 details GraphPFN's in-context-learning (ICL), full fine-tuning (FT), and adapter fine-tuning (AFT) modes alongside GraphPack's fixed training-and-selection workflow. ICL also avoids downstream hyperparameter search, making it a directly relevant comparison. GraphPack completes the measured workflows in 10.0 s and 29.3 s without a pretrained predictor checkpoint, and Q2 below details its gains over the conventional GNN controls.

Methodtolokers-2 (AP)city-roads-M (R2R^2)tolokers-2 runtime (s)city-roads-M runtime (s)
GraphPack58.88 ± 0.6564.11 ± 0.1910.0†29.3†
Plain GAT (incl. tuning)55.21 ± 0.2658.90 ± 0.26765.4703.7
GAT+label (incl. tuning)55.10 ± 0.1860.58 ± 0.16760.7703.1
GCN56.27 ± 0.3158.82 ± 0.2527.24 ± 1.4427.29 ± 3.55
GraphPFN ICL61.29 ± 0.1264.85 ± 0.137.03 ± 0.0416.22 ± 0.02
GraphPFN full FT62.80 ± 0.3967.30 ± 0.21195.44 ± 35.031023.22 ± 232.98
GraphPFN adapter FT61.98 ± 0.5867.53 ± 0.18——

GraphPack and both GAT variants are measured locally on an A800 80 GB. GCN and GraphPFN values are from GraphPFN v4, Tables 1, 7, and 9; their runtimes are the authors' end-to-end measurements at optimal hyperparameters on an A100 80 GB; GraphPFN runtimes include the default ten-member ensemble. “—” means no corresponding runtime is included here. † GraphPack times are complete-run measurements from the original local profiling batch; accuracy and runtime come from separate evaluation and profiling batches.

This gives a direct view of the available workflows: GraphPack's reported times are lower than full FT and higher than ICL, while Q2 details its gains over the conventional GNNs. GraphPack trains without a pretrained predictor checkpoint; supplied node features may themselves be pretrained embeddings.

Peak GPU memory. Separate A800 80 GB profiles cover preprocessing, bank/graph construction, final training, selection, and scoring, excluding preceding search. The table gives the maximum of three process GPU peaks, sampled by NVML every 100 ms.

Methodtolokers-2 GPU peak (GiB)city-roads-M GPU peak (GiB)
GraphPack1.7252.477
SIGN+lab tune-then-ensemble2.8573.689
GAT+label2.0722.594

GraphPack has lower observed GPU peaks than both baselines. These are the separate training profiles from Appendix B.6; the table above reports RL accuracy and workflow costs.

Manuscript revision: Section 4.4 presents the accuracy/cost comparison. Appendix H.2 (Table 28) lists GNN and GraphPFN results with their sources, and Appendix B.6 gives memory profiles. Appendices C.10/H.5 separately identify architecture-only inference timings.

AuthorsOpenReview ↗

Response to Reviewer 9dwR (3/4)

Q1 — Task adaptation on road graphs

GraphPack adapts to a road-prediction task through both learned predictors and ensemble selection. It improves over SIGN+lab and its tune-then-ensemble extension on both city-roads datasets; the new city-roads-M low-label comparison also gives a higher mean than GAT+label.

Evidence that the proposed design benefits road graphs. In Table 3, GraphPack without GBDT achieves mean scores of 77.80 and 76.78 on city-roads-M and city-roads-L, compared with 77.53 and 76.06 for SIGN+lab tune-then-ensemble; the full GraphPack achieves 78.00 and 76.91. Thus, its improvement over this decoupled pipeline is present even without tree candidates. Separately, the controlled neural-only ablation in Table 4 supports the value of sampling views on city-roads-M: sampling both views and training hyperparameters achieves 77.77 ± 0.12, compared with 77.37 ± 0.04 when every member receives all views.

The new city-roads-M comparisons extend this evidence: GraphPack reaches 64.11 ± 0.19 versus GAT+label's 60.58 ± 0.16 on the official RL split. Under strict inner selection, their means are 75.85 ± 0.08 and 75.83 ± 0.26, respectively (Q2/Q4; Appendix H.2/H.4).

Why precomputed views still permit task adaptation. For illustration, consider a bank of multi-hop features B=[X∥SX∥⋯∥SKX]B=[X\|SX\|\cdots\|S^KX], where XX denotes node features and SS a fixed graph propagation operator. Even a linear readout can learn how to combine these views:

BW=∑k=0KSkXWk.BW=\sum_{k=0}^{K}S^kXW_k.

The weight blocks WkW_k are fitted from training labels, so even this simple readout learns which propagation scales to emphasize. GraphPack learns nonlinear MLPs over different view subsets and then selects complementary predictions. The architectural distinction is where task-specific learning occurs: GraphPack learns over a bank of propagation scales and view families, while message-passing GNNs learn transformations between propagation steps. The road-graph comparisons and view ablation show the practical benefit of GraphPack's task adaptation within the decoupled approach.

Manuscript revision: Section 4.2 and Appendix H.3 connect task-adaptive prediction with the road-graph comparisons and structural ablation. Appendix G describes adaptation over the precomputed bank; Appendix H.4 reports the inner-selection experiment.

Q2 — GNN comparisons on the same low-label splits

GraphPack exceeds both locally tuned GAT variants and four published conventional GNN baselines on the same official low-label splits. We use identical RL masks (10%/10%/80% train/validation/test). Each local GAT variant receives its own 15-point validation-selected grid; GraphPack keeps its defaults. These added runs complement the submitted RL results retained in Appendix D.4. W2 reports the local results; the table below gives the published GNN comparisons.

Methodtolokers-2 (AP)city-roads-M (R2R^2)
GraphPack58.88 ± 0.6564.11 ± 0.19
GCN56.27 ± 0.3158.82 ± 0.25
GraphSAGE54.43 ± 0.3459.44 ± 0.27
Plain GAT57.41 ± 0.8559.86 ± 0.20
GT56.98 ± 0.5559.55 ± 0.28

GraphPack reports three runs in this work. The tuned GNN results are from GraphPFN v4, Table 1, with ten runs on the official RL split. These GNNs include residual connections, layer normalization, and MLP blocks; GT uses neighborhood attention.

Relative to the locally tuned plain GAT, GraphPack gains 3.67 AP and 5.21 R² points; relative to GAT+label, it gains 3.78 and 3.53 points. The positive GNN comparison therefore extends across both label-input choices and the four published architectures.

Manuscript revision: Section 4.4 highlights the low-label GNN result. Appendix H.2 (Table 28) reports local and published comparisons, with GraphPFN variants and costs presented separately in W2 above.

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

This paper proposes GraphPack, a one-run ensemble approach for node property prediction that aims to replace expensive per-dataset hyperparameter search. The method precomputes a bank of graph- propagated features, samples different structural views and hyperparameters for 64 MLPs, and trains them together on a single GPU. It then uses validation-based greedy selection to construct an ensemble from the MLPs and additional GBDT candidates. Experiments on several graph benchmarks show competitive performance against tuned GNNs and SIGN-based methods, with lower runtime than the search-and-ensemble approach. The authors also investigate how structural diversity contributes to ensemble performance.

Critical strengths

Simple and practical idea. The idea of replacing expensive hyperparameter search with sampling and ensemble selection is intuitive and practically appealing. Extending this approach to graph-structural choices, rather than only model hyperparameters, is an interesting direction.

Extensive experimental evaluation. The paper includes comparisons against several strong baselines, including tuned GNNs, SIGN, and LightGBM, and evaluates different graph settings, including temporal splits, heterophily, and large-scale graphs. The runtime comparisons are also useful for assessing the method's practical benefits.

Good ablation studies and analysis. The authors provide several controlled experiments to investigate the contributions of structural sampling, hyperparameter diversity, and ensemble composition. In particular, the comparisons with fixed-view and all-view ensembles, along with the analysis of error correlations, provide useful evidence for the benefits of structural diversity.

Critical weaknesses

W1: No supervised feature-selection baseline. The paper includes related comparisons, such as LightGBM on the full bank, SIGN's search over hop and label groups, and the Selected structure and All views arms. However, it does not test supervised selection of informative graph views followed by training a single model. Such a baseline would help determine whether an ensemble is necessary or whether selecting the right views is sufficient.

W2: The contribution of ensemble selection is under-analyzed. All controlled ablation arms use greedy selection, but a uniform average of the trained members is not evaluated. Selected ensembles can be quite small (e.g., 3.3 members on hm-categories; Table 18), making it important to quantify the benefit of selection. W2.1: The distribution of individual member validation scores and the stability of the selected view configurations are also not thoroughly analyzed.

W3: The search comparison is confounded. SIGN+lab differs from GraphPack in its view space, optimizer (AdamW vs. Muon), and lack of GBDT candidates, as the authors acknowledge in Section 4.3 and Appendix C.3. Without GBDTs, GraphPack achieves only 6/3/1 wins/ ties/losses against tune-then-ensemble (Table 3). In the more controlled experiments of Section 5, the advantage over a fixed, selected configuration is modest: 2/4/0 against Tuned ensemble and 0/6/0 against Selected structure (Table 4). While the results support structural sampling, they do not fully isolate its advantage over searching for a good configuration under matched conditions.

W4: The paper empirically connects structural diversity to error decorrelation but does not relate this finding to existing ensemble- diversity theory, such as the ambiguity decomposition or the random subspace method. Additionally, although memory-placement rules are described, peak GPU and host memory usage are not reported. This is particularly relevant given feature banks as large as approximately 193 GB in fp32 for web-fraud, which must be held in host memory.

Questions for authors

Q1: How would GraphPack compare with supervised feature selection on the same bank, such as mutual information or permutation importance, followed by a single MLP?

Q2: Could the authors compare against a search over the same view space and optimizer as GraphPack, with and without GBDT candidates, even on a subset of datasets?

Q3: How does greedy selection compare with simply averaging all 64 MLPs (and the GBDT candidates)?

Q4: How widely do member validation scores vary, and how consistent is the selected ensemble across seeds?

Q5: Could the authors report peak GPU and host memory for GraphPack and representative baselines?

6 comments
AuthorsOpenReview ↗

Response to Reviewer gEoH (1/4)

Thank you for recognizing GraphPack's practical motivation, extensive evaluation, and analysis of structural diversity. Your questions identify three central components: choosing useful views, constructing diverse predictors, and selecting their final combination.

The new controls provide direct evidence for all three. Neural-only GraphPack exceeds supervised-selection single MLPs; greedy selection improves over uniform averaging on identical candidate pools; and different selected compositions produce highly correlated final predictions. We also add matched-space search, theory connections, and measured GPU/host-memory profiles. Below, each weakness is paired with its corresponding question.

Unless otherwise noted, performance is reported as mean ± sample standard deviation over three seeds, with AP, accuracy, and R2R^2 multiplied by 100.

W1 / Q1 — Supervised feature selection followed by a single MLP

Neural-only GraphPack exceeds the stronger supervised-selection baseline by 2.89 points on tolokers-2 and 2.49 points on city-roads-M. Both added selectors use the same propagated feature bank and train a single MLP:

  • MI-selected single MLP: estimate mutual information using training data only, then use validation performance to choose the retained feature fraction from 25/50/75/100%.
  • Group-PI-selected single MLP: jointly permute all columns belonging to each named graph view. Estimate validation-set permutation importance with 10 repeats using a frozen full-bank MLP, then use validation performance to select among single MLPs using the top 1/5/10/15/20 view groups. These candidates are not ensembled.
Methodtolokers-2 (AP)city-roads-M (R2R^2)
Full-bank single MLP (new)59.14 ± 0.4773.87 ± 0.15
MI-selected single MLP (new)60.47 ± 1.0774.44 ± 0.26
Group-PI-selected single MLP (new)59.03 ± 0.3875.31 ± 0.36
GraphPack without GBDT (submitted Table 3)63.36 ± 0.6877.80 ± 0.14
GraphPack full (submitted Table 2)64.42 ± 0.5778.00 ± 0.09

The ensemble benefit remains with neural predictors alone. Group PI improves the city-roads-M single-MLP mean by 1.44 points, while neural-only GraphPack retains the 2.89/2.49-point gains over the stronger selector.

Why selecting views and combining predictors serve different purposes. A single selected subset supplies one predictor with informative inputs. GraphPack trains predictors on different subsets and selects them for their contribution to the ensemble. Figure 3 shows why this matters: members with more distinct view sets have less-correlated residuals, and structural sampling can improve the ensemble even when its best individual member is weaker. The new single-model controls complement that evidence; W2 then isolates how to combine the trained candidates.

Manuscript revision: Section 5 summarizes the result; Appendix C.4 (Table 9) provides the subset grids, permutation repeats, and complete selection protocol.

AuthorsOpenReview ↗

Response to Reviewer gEoH (2/4)

W2 / Q3 — Greedy selection versus uniform averaging

Greedy selection improves the mean on all four tasks, with and without GBDTs, using exactly the same trained candidates and checkpoints. This directly tests the aggregation rule.

To isolate aggregation, each candidate contributes its validation-best checkpoint to a frozen pool. Uniform averaging and greedy selection then operate on the same predictions, using either 64 MLPs or those MLPs plus four GBDTs. Classification probabilities are averaged, and selection uses validation labels only.

DatasetUniform: 64 MLPsGreedy: 64 MLPsUniform: 64 MLPs + 4 GBDTsGreedy: 64 MLPs + 4 GBDTsMean selected size (mixed pool)
tolokers-2 (AP)63.11 ± 0.0663.53 ± 0.1963.53 ± 0.0564.94 ± 0.629.00
city-roads-M (R2R^2)72.25 ± 0.2277.60 ± 0.1072.63 ± 0.1377.87 ± 0.155.67
hm-categories (accuracy)77.45 ± 0.7980.58 ± 0.5578.06 ± 0.6582.51 ± 0.734.33
roman-empire (accuracy)84.39 ± 0.6984.90 ± 0.2784.32 ± 0.5385.13 ± 0.447.00

For the mixed pool, greedy selection improves the mean by 1.41, 5.24, 4.45, and 0.81 points, while retaining only 4.33–9 members on average. Holding candidate training and checkpoints fixed attributes these differences to the aggregation rule.

A small selected ensemble can benefit from a large, diverse pool. Candidate quality varies substantially (W2.1), and adding every candidate gives each equal influence. Greedy selection instead adds a predictor when it improves the current ensemble. It can therefore retain useful complementary predictions while leaving out redundant or unhelpful ones. The gains from identical frozen pools show the value of this choice directly.

Manuscript revision: We have added this control to Appendix E.6 (Table 21), with a summary in Section 5.

W2.1 / Q4 — Member quality and stability across seeds

Different selected compositions produce highly consistent predictions: mean Pearson correlations are 0.991 and 0.994 on the two diagnostic tasks. We analyze member quality, view usage, and final predictions across seeds 0/1/2 on artnet-views and city-roads-M, using the same frozen-checkpoint protocol as above.

Member quality. The following statistics summarize the spread of validation scores. Both datasets use the 100R2100R^2 scale.

DatasetSeedsGlobal validation-score rangeMean within-run medianMean within-run IQR widthMean selected-member median
artnet-views3−54.19 to 63.1755.2711.6061.86
city-roads-M34.14 to 76.7163.2516.7075.42

The wide ranges and IQRs show that the candidate pools contain members of quite different quality. The selected-member medians are higher than the corresponding pool medians.

Selected composition. We compare named-view usage because independently sampled pools contain different configurations. Weighted Jaccard measures overlap in the fractions of selected members using each view; Appendix E.4 gives its definition.

DatasetSelected sizes by seedSelected GBDTs by seedSelected-view weighted Jaccard
artnet-views4 / 6 / 32 / 3 / 20.711
city-roads-M7 / 8 / 22 / 3 / 10.568

Final predictions. After validation-based selection, we compare predictions on the same test nodes to assess consistency across seeds.

DatasetSeed pairsMean pairwise prediction Pearson (range)Mean pairwise prediction Spearman (range)
artnet-views30.991 (0.990–0.992)0.989 (0.987–0.990)
city-roads-M30.994 (0.993–0.995)0.988 (0.985–0.991)

Stable predictions do not require identical selected members. The different view mixtures and ensemble sizes yield highly correlated final predictions. This is consistent with several sampled combinations capturing similar predictive signals, allowing selection to recover a stable predictor from different candidate pools.

Manuscript revision: We have added member-score summary statistics, per-seed composition, the selected-view weighted Jaccard definition, and the prediction-correlation diagnostic to Appendix E.4 (Table 19), with a summary in Section 5.

AuthorsOpenReview ↗

Response to Reviewer gEoH (3/4)

W3 / Q2 — Search using GraphPack's view space and optimizer

We directly tested search over GraphPack's full view and hyperparameter space with Muon, both with and without GBDT candidates. This matches the model choices raised in your question.

The two search routes retain either the validation-best searched MLP or an ensemble selected from all searched configurations. Both routes are evaluated in neural-only and GBDT-augmented settings.

DatasetMethodWithout GBDTWith GBDT
artnet-viewsGraphPack (submitted paper)62.85 ± 0.0563.74 ± 0.19
artnet-viewsTPE validation-best MLP route59.20 ± 0.6563.24 ± 0.34
artnet-viewsTPE searched-configuration selection60.98 ± 0.5062.08 ± 0.49
city-roads-MGraphPack (submitted paper)77.80 ± 0.1478.00 ± 0.09
city-roads-MTPE validation-best MLP route75.62 ± 0.9577.11 ± 0.19
city-roads-MTPE searched-configuration selection77.17 ± 0.3177.02 ± 0.09

GraphPack has the higher mean in every evaluated setting, including selection from all searched configurations and both GBDT settings. This comparison uses the same view space and optimizer and lets the search contribute multiple configurations to an ensemble.

The empirical explanation is that ensemble quality depends on how predictors work together. Figure 3 shows that structural sampling can produce stronger ensembles even with weaker best individual members. The matched-space results support using a broad sampled pool followed by ensemble selection to exploit that complementarity.

The evidence also extends to parallel execution. Appendix H.6 (Table 32) compares eight-worker TPE with packed training under matched resources and a common final selector. On tolokers-2, the packed offline control reaches 64.94 AP in 28.04 s, versus 64.60 AP in 31.32 s for TPE; packed implementations have lower measured GPU peaks on both tasks.

Manuscript revision: Appendix C.6 (Table 10) reports the matched-space comparison; Appendix H.6 (Table 32) adds parallel HPO. Section 4.3 brings the two comparisons together.

W4 — Connection to ensemble-diversity theory

GraphPack creates diversity by varying the neighborhood information seen by each predictor. We relate this design to random-subspace ensembles and use ambiguity decomposition to explain how member quality and diversity jointly affect ensemble error.

Random-subspace connection. In Ho (1998), learners receive different feature subspaces. GraphPack samples overlapping subsets of graph-propagated views: changing operators, hop depths, directions, or label views changes the structural signals a member can use. The shared bank makes these member-specific inputs available without repeating graph propagation during training.

Ambiguity decomposition. For squared loss and nonnegative weights summing to one, the decomposition (Krogh & Vedelsby, 1994) gives

=\sum_i w_i\|f_i-y\|^2 -\sum_i w_i\|f_i-\bar f\|^2, \qquad \bar f=\sum_i w_i f_i.$$ Ensemble error equals weighted member error minus disagreement. The identity explains the benefit of combining accurate predictors whose errors differ, and motivates considering quality and diversity together. It applies to regression and to probability vectors under Brier loss. The structural ablations and error-decorrelation analysis connect this interpretation to graph-view sampling. The fixed-pool experiment in W2 then directly measures the gains from selecting the final combination. Together, the theory connection and controls explain how member quality, diversity, and aggregation contribute to GraphPack. **Manuscript revision:** We have expanded these theoretical connections in Appendix E.1 (Equation 5), with a corresponding statement in Section 5. Section 5 now describes the method as sampling **overlapping subsets of views**.
Reviewer_3yU9OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GraphPack, which samples multiple MLP configurations and precomputed graph-propagation views, trains them together in one packed run, and selects an ensemble as an alternative to extensive hyperparameter search for node property prediction.

Critical strengths
  1. Hyperparameter sensitivity is a significant challenge in graph learning; the search grid can materially affect which model appears best.

  2. Combining diverse propagated graph views is an interesting way to improve robustness across graph tasks.

  3. Experiments cover recent GraphLand .

Critical weaknesses
  1. The work still samples hyperparameters and trains many models, the wall-clock advantage over search may depend on comparing with sequential rather than adequately parallelized HPO. A matched-resource parallel HPO comparison would clarify the gain.

  2. Compare with ensembles using existing graph-view or augmentation techniques, such as DropEdge and subgraph GNNs, under matched compute and parallel resources.

  3. Testing additional emerging domains and tasks, including those highlighted by GraphBench, would strengthen the generality claim.

Questions for authors

please address the weakness with more experiments and ablations as mentioned.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3yU9 (1/3)

Thank you for recognizing the importance of hyperparameter sensitivity and diverse propagated graph views. We have addressed all three requests with new experiments.

GraphPack uses different graph views to train complementary predictors, then selects their combination in one packed run. Each member receives its own mix of propagation operators, hop depths, edge directions, and training-label views. These choices expose members to different neighborhood signals; our structural ablations and error-correlation analysis show how this diversity benefits the ensemble. A shared feature bank and packed training make it practical to explore these choices together. The revision tests this workflow against parallel HPO and graph-augmentation ensembles, and on new graph domains:

  • Parallel-search evidence: competitive prediction and lower measured GPU peaks against eight-worker HPO.
  • Graph-augmentation comparisons: higher means than DropEdge and GraphSAINT ensembles on both added tasks under matched resources and workflow budgets.
  • New-domain evaluation: the same fixed procedure applied to previously unused social and financial graphs.

All new performance and runtime results are mean ± sample standard deviation over seeds 0/1/2. AP, ROC-AUC, and R² are multiplied by 100. Published reference results retain their source protocols.

W1 — Comparison with adequately parallelized hyperparameter search

We completed the requested comparison with eight-worker asynchronous TPE under the same GPU and CPU allocation. GraphPack retains competitive scores and lower measured GPU peaks on both tasks. A packed control with the same final selector as parallel search also provides a direct comparison of the execution workflows.

Matched resources and model choices. Each method uses one exclusive A800 80 GB, a 14-CPU quota, CUDA MPS, and the same RH splits. The primary comparison matches the full view/hyperparameter space, Muon optimizer, 64 MLP candidates, and four GBDTs. TPE runs 64 full trials with eight persistent workers, 16 startup trials, and no pruning; the worker count was chosen by a timing-only pilot. Parallel random search runs GraphPack's sampled configurations in independent workers. The packed offline control and both search methods use one validation-best checkpoint per model and the same final greedy selector, capped at 32 entries. Online GraphPack keeps its original snapshot selection and ensemble early stopping. The no-GBDT rows provide an additional ablation.

DatasetMethodTest AP/R² ×100 ↑Workflow (s)GPU peak (GiB)
tolokers-2GraphPack (online), +4 GBDT64.06 ± 1.0826.79 ± 0.414.34
tolokers-2Packed offline control, +4 GBDT64.94 ± 0.6228.04 ± 2.714.34
tolokers-2Parallel TPE, +4 GBDT64.60 ± 0.9031.32 ± 7.405.59
tolokers-2Parallel random, +4 GBDT64.87 ± 0.6629.59 ± 1.985.59
tolokers-2GraphPack (online), no GBDT63.03 ± 0.6331.03 ± 1.884.34
city-roads-MGraphPack (online), +4 GBDT78.03 ± 0.03107.25 ± 20.156.95
city-roads-MPacked offline control, +4 GBDT77.87 ± 0.15156.86 ± 69.146.95
city-roads-MParallel TPE, +4 GBDT77.99 ± 0.1484.77 ± 11.319.07
city-roads-MParallel random, +4 GBDT77.78 ± 0.1480.91 ± 5.489.07
city-roads-MGraphPack (online), no GBDT77.81 ± 0.05111.93 ± 14.166.95

Workflow time covers loading, preprocessing, bank construction, worker lifecycle, all fits, and selection; Python startup, diagnostic export, and common MPS setup are excluded. GPU peak is the maximum sampled device use across three seeds, including MPS overhead. Appendix H.6 gives the full protocol.

AuthorsOpenReview ↗

Response to Reviewer 3yU9 (2/3)

W1 — Comparison with adequately parallelized hyperparameter search (continued)

What the comparison shows. On tolokers-2, online GraphPack takes 26.79 s, compared with 31.32 s for parallel TPE, with AP means of 64.06 and 64.60. With the common offline selector, the packed control reaches 64.94 AP in 28.04 s, versus TPE's 64.60 AP in 31.32 s. On city-roads-M, online GraphPack has the highest reported mean R², 78.03, while TPE is faster: 84.77 s versus 107.25 s. GraphPack's measured GPU peaks are 4.34/6.95 GiB, compared with 5.59/9.07 GiB for both independent-worker comparators.

Packed execution gives a practical way to train a diverse ensemble even when HPO is parallelized. Here, one run jointly trains the 64 sampled MLP configurations as a single packed module and selects their predictions within that workflow. Members share precomputed graph views and input batches, and their MLP computations run as batched matrix multiplications. The common-selector control connects this implementation to the competitive scores and smaller measured GPU footprint above. The original speedup factors refer to the SIGN+lab search-and-ensemble workflow; this table directly compares parallel HPO.

Manuscript revision: We added this parallel comparison and its protocol to Appendix H.6, with a concise discussion in Section 4.3 distinguishing it from the original SIGN+lab cost comparison.

W2 — Graph-augmentation and subgraph ensemble comparisons

GraphPack achieves the highest mean on both added tasks and exceeds both augmentation techniques in all six matched seed–task comparisons. The new comparison uses ensembles built from DropEdge and GraphSAINT, with matched parallel resources and per-seed workflow budgets.

Matched comparison. Both baselines train eight GNN members under the same A800/14-CPU/MPS allocation, using the RH validation-selected GAT (artnet-exp) and GCN (artnet-views) recipes and the same splits and label inputs as GraphPack. For each dataset and seed, their workflow budget equals the corresponding fresh GraphPack run. Validation selects DropEdge's probability (0.2/0.5) and each method's uniform or greedy aggregation. GraphSAINT uses node sampling with an 80% node budget; Appendix H.7 gives the full setup.

DatasetMetricGraphPackDropEdge ensembleGraphSAINT ensemble
artnet-expAP ↑53.51 ± 0.4852.74 ± 0.3351.15 ± 0.31
artnet-viewsR2R^2 ↑63.75 ± 0.2449.42 ± 1.1228.45 ± 3.46

GraphPack improves over the stronger augmentation baseline by 0.77 AP on artnet-exp and 14.33 R² points on artnet-views. DropEdge and GraphSAINT vary the edges or nodes presented to a GNN. GraphPack varies the structural information available to each predictor: for example, short- and long-range propagation, direction-specific views, and training-label signals. This allows members to learn different combinations of neighborhood information. Table 4 and Figure 3 explain the value of this design: sampling views improves over giving every member all views, and greater structural distance is associated with less-correlated prediction errors. The new comparison demonstrates the effectiveness of this approach against established graph-augmentation ensembles.

Manuscript revision: We added this matched-resource accuracy table and protocol to Appendix H.7 and summarized the result in Section 4.3.

AuthorsOpenReview ↗

Response to Reviewer 3yU9 (3/3)

W3 — Additional domains and tasks beyond the development benchmarks

The fixed GraphPack procedure now covers three held-out graphs from different domains: web, social, and financial. We added GraphBench's BlueSky-quotes engagement prediction and Elliptic financial-transaction classification. On both new graphs, GraphPack improves on the newly trained single-model controls without an additional per-dataset hyperparameter search.

Model defaults and seeds were fixed before observing the added graphs' effectiveness results. The procedure and its configuration distribution transfer to each new graph, where the predictors are fitted using its training data and selected using its validation data. This evaluates the reuse of the training procedure across node-property tasks.

BlueSky-quotes: temporal social engagement. We use the released GraphBench 0.2.0 features and chronological snapshots to predict log1p likes, replies, and reposts. The chronological train/validation/test partitions contain 289,136/286,425/311,272 node-snapshot observations. Propagation stays within each snapshot; label views use training labels only. All three targets and seeds are included. The first three rows are our measurements; the remaining six are published references from GraphBench v5, Table 2. Aggregate MAE averages the three targets within each seed, then reports the mean and sample SD across seeds.

MethodAggregate MAE ↓
GraphPack0.7387 ± 0.0023
Fixed full-bank single MLP0.8008 ± 0.0021
Raw-feature LightGBM0.7882 ± 0.0001
DeepSets0.810 ± 0.005
MLP0.784 ± 0.001
GNN0.768 ± 0.002
GIN+0.737 ± 0.014
GatedGCN+0.734 ± 0.010
GCN+0.781 ± 0.021

GraphPack reduces aggregate MAE from 0.8008 for the full-bank single MLP and 0.7882 for LightGBM to 0.7387. Its R² scores for likes/replies/reposts are 23.77 ± 0.31 / 20.54 ± 0.37 / 23.94 ± 0.45. All three target-wise R², MAE, and Spearman scores improve on both controls. Its aggregate MAE is also below the published means of DeepSets, MLP, GNN, and GCN+, supporting the competitiveness of the fixed procedure on this new temporal social graph.

Elliptic: financial-transaction classification. On Elliptic, we use chronological fit/validation/test periods 1–29/30–34/35–49 with the same fixed model defaults. The released features include neighborhood aggregates. Alongside the three new measurements, we include seven independent literature references from GADBench, Tables 4 and 13. GADBench's AUPRC is average precision; its values are point estimates from tuned methods under its source protocol.

MethodSourceAP ↑ROC-AUC ↑
GraphPackThis work79.32 ± 0.2992.49 ± 0.89
Fixed full-bank single MLPThis work73.02 ± 3.3790.02 ± 0.84
LightGBM on released featuresThis work78.92 ± 0.0092.40 ± 0.00
GCNGADBench25.3781.69
GraphSAGEGADBench57.8287.59
BGNNGADBench62.0387.69
XGBoostGADBench76.9389.91
XGB-GraphGADBench77.7891.78
Random ForestGADBench78.4291.68
RF-GraphGADBench78.8693.21

GraphPack achieves 79.32 AP, compared with 73.02 for the full-bank single MLP and 78.92 for LightGBM. Its ROC-AUC mean is also above both controls. The AP mean is numerically above the seven published reference values shown here, providing additional context for the competitiveness of the fixed procedure in this financial domain.

Fixed sampling rules still allow the predictor to adapt to each graph. The bank offers multiple structural signals, predictors learn from the new graph's labels, and validation selects their combination. Figure 4 illustrates this adaptation: the selected view mix changes across graphs under the same sampling rules. The new social-regression and financial-classification results demonstrate the usefulness of reusing this procedure across domains.

Manuscript revision: Sections 4.1 and 4.4 and Appendix H now distinguish development benchmarks from held-out transfer evidence. Appendix H.1 reports the BlueSky and Elliptic protocols and results in Tables 25–27, including all three BlueSky targets and their reported metrics.

Thank you again for your constructive suggestions. We hope the new comparisons and explanations address your concerns.

Thank you again for your time and thoughtful review. We hope our response has addressed your concerns. Please let us know if you have any further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global response

Thank you for the careful reviews and constructive suggestions. GraphPack trains complementary predictors from different graph views and selects their combination in one packed run.

The reviews recognize the motivation and value of this design: 3yU9 highlights hyperparameter sensitivity and diverse propagated views; gEoH recognizes the practical approach and controlled analyses; and 9dwR emphasizes the practical contribution of effective, structurally diverse graph ensembles.

Key additions

  • Practical comparisons. We add same-space search and eight-worker parallel HPO. GraphPack retains competitive prediction and lower measured GPU peaks in the parallel comparison, and exceeds DropEdge and GraphSAINT ensembles on both added tasks.
  • Evidence for the design. Neural-only GraphPack exceeds the stronger supervised-selection single-MLP baseline by 2.89/2.49 points. Greedy selection improves all four fixed-pool tasks, with and without GBDTs. Member-stability diagnostics and ensemble-theory connections further explain the role of diversity and selection.
  • Transfer and label budgets. Fixed defaults outperform both new single-model controls on BlueSky-quotes and Elliptic, extending evaluation to three held-out graph domains. New low-label tests show gains over conventional GNNs on two tasks; selection restricted to the original training-label budget retains gains over plain GAT and a full-bank MLP.

We also bring the GraphPFN accuracy/cost comparison into the main text and report GPU/host-memory profiles. Sections 4 and 5 present the key findings, with full protocols and results in the appendix.

The individual responses address every listed weakness and question. Thank you again for helping us improve the paper; we hope the new evidence and explanations resolve your concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.