acceptodds
Under review as a conference paper at ICLR 2027

Effective Diameter: Over-Smoothing in the Context of Limited Problem Radii

Abstract

Over-smoothing is widely considered to be one of the main reasons why Graph Neural Networks (GNNs) achieve their best performance with only a few message passing layers. In this paper, we critically examine this assumption on real-world literature benchmarks. Over-smoothing is theoretically well understood and effectively mitigated by established techniques. In experiments on common node classification benchmarks, we show that these mitigation techniques largely stabilise GNN performance for depths of up to 64 layers, but deep models still fail to significantly outperform shallow ones. As an alternative explanation for shallow optimal GNN models, we propose the *effective diameter*, a data-based statistic free of any model assumptions that estimates the size of the task-relevant receptive field. It builds on the shell *separability gap*, which aims to measure the feature-based class discriminative power of a -shell. On all investigated datasets, the separability gap decays within a few hops, and the best-performing models rarely lie beyond the estimated effective diameter. This suggests that the small optimal depth of GNNs is caused by the lack of task-relevant information in large receptive fields rather than by over-smoothing.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Over-smoothing, Receptive Field

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_m5r6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper questions the common belief that over-smoothing is why deep GNNs do not beat shallow ones. The authors show that standard mitigation techniques do prevent over-smoothing up to 64 layers, yet deep models still do not significantly outperform shallow ones on 6 node-classification benchmarks. Alternatively, they propose the "effective diameter": a model-free statistic built on the shell separability gap Δ(k)\Delta(k), which measures whether the k-th shell of a node still carries feature information about the node's class. They find that Δ(k)\Delta(k) decays within a few hops, and that the best-performing model depth rarely exceeds the estimated effective diameter. The conclusion is that the small optimal depth comes from the lack of task-relevant information in large receptive fields, not from over-smoothing.

Critical strengths
  1. Well-motivated question with clean logic. The counterfactual argument is convincing: if over-smoothing were the main bottleneck, fixing it should let deep GNNs win — they still don't, so the bottleneck likely lies elsewhere.
  2. Simple and cheap estimator. Δ(k)\Delta(k) and the effective diameter are model-free, easy to compute (O(Nd)O(Nd) via class-mean decompositions).
Critical weaknesses

W1 — The core statistics are heuristic and the threshold τ\tau.

  1. Δ(k)\Delta(k) is essentially a Fisher-style between/within second-moment ratio; the O(N²d) → O(Nd) simplification is not novel. The only justification given is a null-baseline argument (random relabeling drives the gap to zero), which is a necessary condition, not evidence that a large gap implies downstream-useful information. No quantitative link between Δ(k)\Delta(k) and classification error is provided.
  2. The effective diameter depends on a hand-picked τ\tau (0.005, 0.01 for Photo) with no selection rule and no sensitivity analysis: smaller τ gives a larger diameter, and the paper's only validation: "best depth ≤ effective diameter", a one-sided inequality that becomes trivially true as τ\tau shrinks. Also, on Cora itself Δ(k)\Delta(k) rises again after k=4k = 4, which sits oddly with the "stagnation" story.
  3. "Deep models don't win" could also reflect an optimization bias toward shallow models rather than missing information (deep models are harder to train than shallow ones); no ablation separates the two.

W2 — The empirical study does not look reliable. Baseline numbers are inconsistently off from the literature in both directions: GCN on Cora gets 74.1% where the standard planetoid result is >80%. This inconsistency suggests implementation or protocol problems. The hyperparameter searching space is small, no weight decay and learning rate tuning. GCN on Cora at 4 layers collapses to 36.1 ± 13.4, which looks like a training failure rather than over-smoothing. On Roman-Empire, standard deviations of 10–15pp on the best models (GCNII, G2) are far larger than the depth effects being discussed (e.g., 55.1 vs 57.7). Appendix B claims OGBN-arXiv is included, but no arXiv results appear anywhere in the paper.

W3 — The statistic only covers the raw-feature, but the claim is stated generally. Δ(k)\Delta(k) is computed on input features only. GNNs also do feature learning, and label/structure propagation alone can carry task-relevant signal even when features are uninformative (e.g., label propagation on feature-weak graphs). None of the 6 benchmarks can disentangle the feature channel from the label/structure channel, so "large receptive fields lack task-relevant information" is only shown for one channel, while the claim and conclusion state it without this qualifier.

W4 — The estimator cannot handle non-monotone informativeness across hops. The effective diameter is defined as the first k where Δ(k)≤τ\Delta(k) ≤ \tau, so it has no mechanism for information that reappears at larger hops, e.g., even-hop-dominant structure on heterophilic graphs (as in EvenNet-style arguments). The late spike of Δ(k)\Delta(k) on Amazon-Computers (Figure 6) is acknowledged in one sentence but not incorporated into the estimator. Heterophilic datasets of this type (e.g., Pokec) are neither tested nor discussed.

Questions for authors

Q1 (for W1). How exactly is τ\tau chosen, and how do the estimated effective diameters and the "best depth ≤ effective diameter" result change if τ\tau varies over, say, [1e-3, 1e-2]? Can you rule out that shallow optima come from optimization preferences rather than missing information?

Q2 (for W2). Why do the GCN baselines deviate from the literature? Can you reveal more implementation details? Where are the OGBN-arXiv results announced in Appendix B?

Q3 (for W3). Does your conclusion hold on datasets where raw features are weak but structural/label information is strong?

Q4 (for W4). How would the estimator behave if informativeness is non-monotone in k (uninformative odd hops, informative even hops), like Pokec dataset?

4 comments
AuthorsOpenReview ↗

Rebuttal 1/2

We would like to sincerely thank you for your review. Below we respond to your weaknesses and questions jointly.

Quantitative link between Δ(k)\Delta(k) and classification error (W1.1): We analysed the correlation between the predictive performance of the models GCN, GATv2 and GraphSAGE and the separability gap. We neglected the models with mitigation techniques here, as these architectures stabilise the performance. The results in Table 1 show a positive correlation.

Table 1: Correlation analysis of separability gap Δ\Delta and predictive performance.

DatasetnPearson rrSpearman ρ\rho
Cora90.87120.8960
CiteSeer90.93520.8960
PubMed90.44610.4216
Roman-empire110.84450.9419
Photo60.32270.2928
Computers40.78440.8944

Choice of τ\tau (W1.2 & Q1): We chose τ\tau by jointly observing the curves of the separability gap Δ\Delta and identifying a value of τ\tau to visually fit the stagnation of Δ\Delta. In Table 1, we report the effective diameter for variations of τ\tau, in comparison to the graph diameter and the best model depth. In all cases, the effective diameter remains more informative for the best number of GNN message passing layers than the graph diameter.

Sensitivity of τ\tau (Q1): In Table 2, we report the effective diameter for variations of τ\tau, in comparison to the graph diameter and the best model depth. In cases, where the threshold τ\tau is never reached, we set the effective diameter to the radius with the smallest value of Δ\Delta. In all cases, the effective diameter remains more informative for the best number of message passings with a GNN than the graph diameter.

Table 2: Sensitivity analysis for the effective diameter for different values of τ\tau.

Datasetτ=0.01\tau = 0.01τ=0.005\tau = 0.005τ=0.001\tau = 0.001DiameterDepth$^\star$
Cora447192 (gcn)
CiteSeer5712282 (gcn)
PubMed345182 (gcn)
Photo444112 (gcnii)
Computers335102 (gcnii)
Roman-empire23968244 (gcnii)

Stagnation of Δ\Delta on Cora (W1.2): The increase of Δ\Delta after stagnation for long paths on Cora and other datasets can be attributed to the small number of paths of that length (as visible in the third row of plots in Figure 2 of our manuscript). The growth is not significant, as the variance is very high for these measurements. The same is true for the small decrease after the plateau at k>10k>10 on Roman-empire.

Optimisation preference (W1.3 & Q1): We try to rule out an optimisation preference by searching for the optimal hyperparameters for each model depth individually. The stable results for the GCNII and G2 indicate that there is no optimisation problem with deep models in our experimental suite.

Statistical reliability (W2 & Q2): We investigated the results for the planetoid datasets and found the cause of the deviation from the literature. Initially, we conducted a cross-validation for the hyperparameter search and early stopping during testing on the combined training and validation set. The implementation of our GCN also deviated from the implementation by Kipf and Welling [1], as we did not apply dropout and row-normalisation to the raw input features before. We reran all hyperparameter optimisations and evaluations for all three citation networks, also including the learning rate as a hyperparameter and updated the tables and figures in our manuscript. The GCNII now has lower accuracy across all depths, possibly due to fixing the validation set instead of using cross-validation. The performance of the GCN has increased due to the exact replication of the implementation based on Kipf and Welling [1]. We also found a mistake in our implementation of the Errica protocol, which potentially caused the high variance on the Roman-empire dataset. We updated these results as well. The accuracy with a GCNII is much higher now, but the best depth remains low with 4 layers of message passing. Thank you for making us aware of the error of mentioning OGBN-arXiv; we removed it in the revised manuscript.

[1] Kipf, Thomas N., and Max Welling. “Semi-Supervised Classification with Graph Convolutional Networks.” International Conference on Learning Representations (ICLR), 2017.

AuthorsOpenReview ↗

Rebuttal 2/2

Weak features (W3 & Q3): We investigated the robustness of our method against weak feature signals with an ablation study on Cora. We perturbed the raw input features X∈RN×dX \in \mathbb{R}^{N \times d} with Gaussian noise scaled per feature by its empirical standard deviation sjs_j: x~ij=xij+σsjεij\tilde{x}{ij} = x{ij} + \sigma s_j \varepsilon_{ij} with εij∼N(0,1)\varepsilon_{ij} \sim \mathcal{N}(0,1). The noise level σ\sigma is therefore the ratio of noise to signal standard deviation for every feature. The noise is sampled once before training and stays fixed throughout. We ran the full grid search and evaluation of a GCN and ran the effective diameter estimator. We report the results in Table 3. The effective diameter is not affected by the noise. The optimal GCN depth increases with the level of noise. These results reflect the limitation of our method to graphs with task-relevant features, as we discuss in Section 5.4 of the manuscript. At low signal to noise ratios in the features, the features gradually stop being task-relevant.

Table 3: GCN performance and effective diameter on Cora with different levels of noise applied to the features.

Noise Levelk=1k=2k=3k=4k=5k=6k=7k=8k=9k=10Efffective DiameterBest depthDiameter
σ=0.0\sigma=0.0Rel. gap Δ\Delta(%)4.893.101.130.500.270.170.150.190.290.394219
Accuracy (%)74.78 ± 0.6180.27 ± 0.9679.64 ± 1.0076.84 ± 5.6080.35 ± 0.9178.48 ± 1.5378.29 ± 1.8075.57 ± 2.8275.63 ± 2.4873.58 ± 2.92
σ=0.5\sigma=0.5Rel. gap Δ\Delta(%)4.522.861.060.470.250.160.140.180.270.364219
Accuracy (%)70.70 ± 0.6779.34 ± 0.5978.44 ± 1.0279.47 ± 0.6676.36 ± 5.5578.40 ± 1.8876.68 ± 3.5876.94 ± 1.3975.40 ± 1.4674.14 ± 2.27
σ=2.0\sigma=2.0Rel. gap Δ\Delta(%)3.512.230.870.380.220.160.120.140.210.273719
Accuracy (%)49.67 ± 0.6868.52 ± 1.0172.39 ± 0.7773.27 ± 0.8872.52 ± 1.2669.83 ± 7.4674.61 ± 1.1272.64 ± 1.8972.49 ± 2.3667.68 ± 3.49
σ=8.0\sigma=8.0Rel. gap Δ\Delta(%)3.212.060.810.350.210.160.120.120.180.243819
Accuracy (%)37.61 ± 0.6552.55 ± 1.2760.88 ± 1.4865.57 ± 1.3767.30 ± 1.1666.75 ± 2.6166.71 ± 1.8968.82 ± 1.3968.65 ± 2.0768.27 ± 1.54

Non-monotone informativeness (W4 & Q4): Non-monotone informativeness in general can be observed in the new results for the City Network datasets [2], which also include long-range dependencies. In Figure 7 of the revised manuscript, we can observe that the separability gap first grows, then shrinks and finally saturates for all four graphs. The effective diameter is in all cases much closer to the optimal model depth of 16 [2] than the graph diameter. We also applied our estimator to the pokec datasets. From the results in Table 4, we can not observe non-monotone informativeness and therefore chose to focus this discussion on the four city networks.

Table 4: Relative separability gap Δ\Delta on the dataset Pokec.

Datasetk=1k=2k=3k=4k=5k=6k=7k=8k=9k=10Horizon
PokecRel. gap Δ\Delta(%)0.030.010.000.000.00-0.000.00------1

[2] Liang et al. “Towards Quantifying Long-Range Interactions in Graph Machine Learning: A Large Graph Dataset and a Measurement.” International Conference on Learning Representations (ICLR), 2026.

Reviewer_m5r6OpenReview ↗

Response to rebuttal

We thank the authors for the detailed rebuttal and the substantial amount of new experiments. My concerns are partially resolved. Please see the issues below.

(W1.3) The optimization-preference concern also stands: a per-depth hyperparameter search does not separate "no task-relevant information at this depth" from "this depth is harder to train", and no ablation distinguishes the two.

(W3) The feature-noise ablation undermines rather than supports the main claim. At sigma = 8, the effective diameter is 3 while the best GCN depth is 8: the paper's central inequality "best depth <= effective diameter" is violated by the authors' own table. In addition, the label/structure channel (Q3) was not tested: adding Gaussian noise to features is not the same as a feature-weak graph in which label propagation carries the signal.

(W4) My question was not directly answered. On Pokec the estimator reports Delta close to 0 with horizon 1, but the actual optimal depth on Pokec is not reported, so whether this constitutes a counterexample remains open.

Given the above, I maintain my rating of 2.

We thank you for your detailed additional comments.

(W1.3) We argue that suitable hyperparameter optimisation and adequate model selection do separate the relevant depth interpretation from optimisation problems: On the datasets PubMed and Roman-empire, the GCNII achieves the overall optimal performance at 64 and 32 layers, respectively. This shows that the model is very unlikely to have training problems. We would be grateful for further details on why this evidence is not sufficient and what ablation study could resolve this issue.

(W3) We included the limitation of our method to graphs with informative features in our manuscript. We are happy to refine this section if it does not cover the case that you are discussing.

(W4) We agree that the Pokec dataset does not help to clarify the behaviour of our estimator if the informativeness is non-monotone, as we cannot observe non-monotone informativeness in Pokec. The datasets of the City Network, however, show a non-monotone separability curve and demonstrate the practical usefulness of our method.

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

This paper challenges the prevalent assumption that over-smoothing is the primary factor limiting the depth and performance of Graph Neural Networks (GNNs). Through empirical analysis across six node classification benchmarks using various backbones and mitigation techniques, the authors demonstrate that while mitigation techniques successfully prevent representation to collapse, deep models still fail to outperform shallow ones. As an alternative explanation, the authors propose the effective diameter. It is a model-agnostic statistic derived from the shell separability gap, which measures the relative difference between within-class and between-class feature distances at shell k along sampled shortest paths. Evaluation results show that the separability gap decays rapidly within a few hops, which demonstrates that the performance ceiling is primarily driven by the lack of task-relevant information at larger receptive fields rather than over-smoothing.

Critical strengths

S1: This paper conducts comprehensive empirical analysis, which convincingly demonstrates that resolving over-smoothing does not unlock superior performance in public gaph learning benchmarks for deep GNNs.

S2: The proposed shell separability gap and effective diameter quantify task-relevant receptive fields directly from data and is model-agnostic.

S3: The experimental evaluation is comprehensive, where the authors track multiple over-smoothing metrics across various GNN backbones and mitigation techniques.

Critical weaknesses

W1: The identification of the effective diameter relies on visually chosen threshold values τ\tau.

W2: The proposed metric relies strictly on raw pairwise Euclidean distances across node features, which cannot capture complex non-linear combinations or structural graph motifs.

W3: The experimental evaluation focuses exclusively on traditional citation and co-purchase graphs. Validation on established long-range benchmarks is missing.

Questions for authors

Q1: How sensitive is the calculated effective diameter to variations in τ\tau?

Q2: What does the shell separability gap look like on long-range benchmarks? Does the proposed metric accurately reflect a deeper effective diameter on datasets where depth has been proven beneficial?

2 comments

We want to warmly thank you for your considered review. Below, we address your identified weaknesses and your question in turn.

Choice of τ\tau (W1): We discuss the visual judgement for the choice of the threshold τ\tau as a limitation in Section 5.4. In our submission and the experiments during the discussion phase, we identified the threshold τ=0.005\tau = 0.005 to be remarkably informative for the investigated datasets. Also, it seems that our method is relatively robust to the choice of this visually chosen threshold, as the sensitivity analysis in our response to Q1 shows.

Structural graph motifs (W2): Under the assumption that the node features are task-informative, the general graph structure is captured in our method through the path sampling approach. Statistics that capture specific graph motifs already exist in the literature [1,2,3]. However, these are purely structural. We view the combination of structural and feature signals in our measure as a strength. Nevertheless, we discuss the importance of task-relevant features as a limitation in our manuscript in Section 5.4.

Non-linear feature interactions (W2): In Section 5.4, we discuss that using the Euclidean distance is the result of a trade-off between exactness and computational efficiency. We intended our metric to be cheap and free of model assumptions, while still being more informative than the graph diameter when it comes to choosing an adequate model depth. On all datasets, the effective diameter is much closer to the optimal depth than the graph diameter and has a much shorter runtime. We added a sentence in Section 5.4 to clarify that the effective diameter can only give an indication of the optimal model depth, but cannot replace careful hyperparameter tuning.

Long-Range Benchmark (W3 & Q2): We computed the gap Δ\Delta and the effective diameter for all four graphs of the City Network Benchmark [4]. This benchmark is based on a real-world graph, while the labels for the prediction task are generated synthetically to require information for the 16-hop neighbourhood of a node. The effective diameter with the same threshold τ=0.01\tau = 0.01, as used in the paper for the Photo dataset, is reported in Table 1 below. The effective diameter is close to the optimal depth of a GCN according to [4], while the graph diameter is orders of magnitude larger, which indicates that our method generalises to tasks that require long-range dependencies. We added the visualisations of the separability curves in Appendix E of the revised manuscript.

Table 1: Effective Diameter on City Network

DatasetOptimal GCN Depth [4]Effective DiameterDiameter
Paris1620183
London1617501
LA1619382
Shanghai1618224

Sensitivity of τ\tau (Q1): In Table 2, we report the effective diameter for variations of τ\tau, in comparison to the graph diameter and the best model depth. In cases, where the threshold τ\tau is never reached, we set the effective diameter to the radius with the smallest value of Δ\Delta. In all cases, the effective diameter remains more informative for the best number of message passing layers in a GNN than the graph diameter.

Table 2: Sensitivity analysis for the effective diameter for different values of τ\tau.

Datasetτ=0.01\tau = 0.01τ=0.005\tau = 0.005τ=0.001\tau = 0.001DiameterBest Depth
Cora447192 (gcn)
CiteSeer5712282 (gcn)
PubMed345182 (gcn)
Photo444112 (gcnii)
Computers335102 (gcnii)
Roman-empire23968244 (gcnii)
Paris20232318316 [4]
London1721>= 2450116 [4]
LA1921>= 2438216 [4]
Shanghai18>= 24>= 2422416 [4]

[1] Watts et al. 1998. “Collective Dynamics of ‘Small-World’ Networks.” Nature 393 (6684): 440–442.

[2] Milo et al. 2004. “Superfamilies of Evolved and Designed Networks.” Science 303 (5663): 1538–1542.

[3] Shervashidze et al. 2009. “Efficient Graphlet Kernels for Large Graph Comparison.” Proceedings of Machine Learning Research (PMLR): 488-495.

[4] Liang et al. 2026. “Towards Quantifying Long-Range Interactions in Graph Machine Learning: A Large Graph Dataset and a Measurement.” International Conference on Learning Representations (ICLR).

Thank you for your detailed response. I will keep my score at 3.

Reviewer_uQUhOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

This paper challenges the common assumption that over-smoothing is the primary reason why deep GNNs often fail to outperform shallow models. To provide an alternative explanation, the authors introduce Effective Diameter, a model-agnostic, data-driven metric based on the Shell Separability Gap, which measures the class-discriminative information available at different neighborhood distances.

Critical strengths

I particularly appreciate that the authors do not simply propose another conventional method for mitigating over-smoothing, but instead shift the research perspective toward whether the task itself requires long-range information. I find this perspective highly valuable. Moreover, compared with approaches that require training GNNs of different depths, the proposed path-sampling and feature-distance analysis offers the potential to identify promising receptive field sizes in advance.

Critical weaknesses
  1. The authors argue that the limited advantages of deep GNNs cannot be attributed solely to over-smoothing, and provide experimental evidence in Table 1 to support this claim. However, could the authors further rule out other confounding factors or establish a causal relationship through additional analysis? This would provide stronger support for their argument.

  2. A separability gap of zero does not necessarily imply that long-range information has no predictive value. The relationship between the separability gap and the predictive utility of long-range information also requires further experimental validation or theoretical justification from a causal perspective.

  3. The current experiments mainly rely on existing node classification benchmarks, which may inherently depend more heavily on local information. This raises concerns about whether the conclusions can generalize to tasks that require long-range dependencies.

  4. The Introduction states that deep GNNs fail to significantly outperform shallow ones, but does not sufficiently explain why this motivates investigating task-relevant receptive fields rather than other possible factors. If the main hypothesis is that tasks do not require large receptive fields, the authors should clarify this motivation. Moreover, the paper places too much emphasis on over-smoothing, including in the title, while its main contribution is the Effective Diameter metric. The authors should reconsider the paper's motivation and overall narrative.

Questions for authors

See the weakness.

4 comments

We thank you very much for your review. We now respond to your identified weaknesses point-by-point.

Confounding factors (W1): We tried to rule out several confounding factors by conducting a separate hyperparameter search for each model depth and each model. We included the hidden dimension and dropout in the grid search and used early stopping to mitigate overfitting. With pairnorm and GCNII, which includes skip-connections, we tested two model variants which provide improved gradient flows and stabilise training. We would be happy to investigate further confounding factors if you name them.

Validation of predictive performance and separability gap (W2): We analysed the correlation between the predictive performance of the models GCN, GATv2 and GraphSAGE and the separability gap. We neglected the models with mitigation techniques here, as these architectures stabilise the performance. The results in Table 1 show a positive correlation.

Table 1: Correlation analysis of separability gap Δ\Delta and predictive performance.

DatasetnnPearson rrSpearman ρ\rho
Cora90.87120.8960
CiteSeer90.93520.8960
PubMed90.44610.4216
Roman-empire110.84450.9419
Photo60.32270.2928
Computers40.78440.8944

Long-Range Benchmarks (W3): We computed the gap Δ\Delta and the effective diameter for all four graphs of the City Network Benchmark [1]. This benchmark is based on real-world graphs, while the labels for the prediction task are generated synthetically to require information for the 16-hop neighbourhood of a node. The effective diameter with the same threshold τ=0.01\tau = 0.01, as used in the paper for the Photo dataset, is reported in Table 2 below. The effective diameter is close to the optimal depth of a GCN according to [1], while the graph diameter is orders of magnitude larger, which indicates that our method generalises to tasks that require long-range dependencies. We added the visualisations of the separability curves in Appendix E of the revised manuscript.

Table 2: Effective Diameter on City Network

DatasetOptimal GCN Depth [1]Effective DiameterDiameter
Paris1620183
London1617501
LA1619382
Shanghai1618224

Motivation (W4): We agree with the reviewer that our manuscript can benefit from more clarity about the motivation of our method. We tried to improve the explanations of our motivation in the introduction in the revised manuscript. We view the analysis of over-smoothing and its findings as the fundamental motivation for our proposed method to estimate the effective diameter. A common narrative in the literature is that over-smoothing prevents better GNN performance because it limits the depth of GNNs. We aim to break this narrative and give an alternative explanation: most tasks have a small problem radius. With the effective diameter, we can quantify this alternative hypothesis. With the title, we want to emphasise the intention. We welcome further comments on our explanations.

[1] Liang et al. “Towards Quantifying Long-Range Interactions in Graph Machine Learning: A Large Graph Dataset and a Measurement.” International Conference on Learning Representations (ICLR), 2026.

Thank you for the clarification. While the separate hyperparameter searches for each model and depth help ensure a fairer comparison, I do not find this sufficient to address my concern about confounding factors. My concern is therefore about the interpretation of the results, rather than whether the models were adequately tuned. More targeted controlled experiments, theoretical justification, or further analysis would be needed to strengthen the proposed explanation.

Thank you for your quick response. We would be happy to add further experiments if you could provide some actionable, more specific details on which confounding factors or further analyses you have in mind.

For example, Table 1 shows that mitigating over-smoothing does not always enable deep models to outperform shallow baselines. However, it does not distinguish between long-range information lacking predictive value and the models failing to effectively exploit that information. Could the authors further validate their explanation through controlled experiments that vary the task-relevant information distance, or by examining whether distant neighborhoods provide additional predictive value beyond local information?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.