acceptodds
Under review as a conference paper at ICLR 2027

Graph Convolution with Degree-Potential and Spectral Calibration

Abstract

Graph neural networks (GNNs) are widely used for node classification, aggregating neighborhood information through rules in which degree influences message weighting. Degree-aware methods have extended these rules through adaptive aggregation and structural augmentation. We investigate how relative source–receiver degrees can guide shared and local message calibration on the existing sparse graph. On PubMed, matching receiver-degree distributions reveals larger GCN–mean message changes in neighborhoods with greater degree contrast. This motivates degree-potential calibrated graph convolution (DPC-GCN), which combines shared and centered receiver-dependent corrections. To examine additional spectral flexibility after degree calibration, we introduce signed spectral calibration (SSC), yielding DPC-GCN-S. One shared scalar per layer learns the direction and strength of state adjustment, allowing preservation, smoothing, or sharpening. Our analysis explains the separation of shared and local corrections and shows that identical label spectral energies can admit opposite beneficial local calibration directions. Across 16 benchmarks, including three large-scale graphs, DPC-GCN matches or exceeds the best baseline mean on all eight homophilous datasets; DPC-GCN-S ranks first on four of five heterophilous datasets among the evaluated methods. An anonymized implementation is available at https://anonymous.4open.science/r/DPCGCN-4ED8.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph neural networks, node classification, degree-aware propagation, spectral filtering, heterophily

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 2 reviews, above 68% of the others.
Reviewer_xczWOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces degree-potential calibration (DPC), which adjusts sparse GCN propagation as PDPC=Pg+αΔ+λDiag⁡(t)ΔP_{\mathrm{DPC}}=P_g+\alpha\Delta+\lambda\operatorname{Diag}(t)\Delta, where Δ=Pm−Pg\Delta=P_m-P_g and tt is a centered degree-discrepancy coordinate. An optional signed spectral calibration (SSC) applies I−τLgI-\tau L_g, with one bounded learned scalar per layer, to smooth or sharpen the resulting state. The analysis establishes separation of shared and local corrections and constructs labelings with identical spectral energies and initial margin distributions but opposite locally beneficial calibration signs. Across 16 benchmarks, DPC matches or exceeds the highest baseline mean on all eight homophilous datasets, while DPC-GCN-S leads on four of five heterophilous datasets. Ablations examine the coordinate construction, sign, parameter granularity, and resource costs.

Critical strengths
  1. A precise, economical construction with substantive analysis. The centered coordinate makes the local correction orthogonal to the shared mixing direction in the Frobenius geometry, giving a clean interpretation of the additional parameter. The spectral response of SSC is explicit. More unusually, Proposition 3 controls both spectral energy and the initial margin distribution while reversing the beneficial risk derivative. Its explicit C18C_{18} construction is internally consistent under an independent numerical check at σ=1\sigma=1. This is a useful caution against choosing a smoothing/sharpening direction solely from aggregate spectral statistics.

  2. Broad evaluation with relevant design controls. The experiments include tuned local GNNs and modern graph Transformers across graph types and scales. Appendices B.2–B.3 compare centered, uncentered, shuffled, sign-constrained, feature-wise, node-wise, and fixed-propagation variants. The paper also reports efficiency at the selected operating points and the accuracy deficit on ogbn-products. These controls make the evidence more informative than an aggregate leaderboard alone.

Critical weaknesses
  1. The empirical necessity of the degree-dependent local correction is limited. Proposition 1 establishes geometric non-redundancy, which does not itself establish improved prediction. In Table 5, removing DPC while retaining SSC gives exactly the same displayed mean on Roman-Empire and Questions, and costs only 0.150.15 and 0.100.10 points on Minesweeper and Tolokers. Table 6 also shows that graph-wide or shuffled-centered controls are competitive on several tasks. Thus the strongest empirical case for the complete construction is concentrated on a subset of datasets. This limits how much the results establish about the distinctive degree-dependent component, even though the compact combined method remains useful.

  2. Small leaderboard gains need stronger control of model-selection uncertainty. Several improvements over the strongest baselines are fractions of a percentage point. Appendix C.3 selects degree-coefficient initializations and SSC learning-rate multipliers, while Appendix C.4 uses different candidate/configuration policies for different baselines. The paper is transparent about these choices, but does not give comparable candidate counts or paired uncertainty for the decisive accuracy differences. The resource comparisons at separately selected widths/depths are useful operating-point comparisons, rather than an isolated measurement of the calibration operator's overhead. These issues limit the strength of a general superiority claim; they do not invalidate the reported ablations or the spectral derivations.

Questions for authors
  1. Evidence request: the value of the local degree coordinate. Using the same run/split pairs, what are the accuracy differences and uncertainty intervals between full DPC, graph-wide mixing, and shuffled-centered coordinates in Table 6, particularly on Computer, WikiCS, and Minesweeper? This would clarify whether the reported advantage of the topology-specific local correction is consistent across runs.

  2. Clarification: the contribution of DPC beyond SSC. In Table 5, retaining SSC while removing DPC gives the same displayed mean on Roman-Empire and Questions and only slightly lower means on Minesweeper and Tolokers. Which empirical results most strongly support the incremental benefit of DPC once SSC is present? Please distinguish evidence for the degree-calibration component from evidence for the combined architecture.

  3. Clarification: comparable model-selection effort. What are the effective numbers of configurations evaluated for DPC and the strongest tuned GNN baselines, including coefficient initializations and SSC learning-rate multipliers? Since several leaderboard margins are small, please explain whether the conclusions remain supported after accounting for differences in model-selection effort.

  4. Evidence request: efficiency attributable to calibration. The existing resource comparisons use each method's selected configuration. At matched hidden width, depth, and training protocol, what additional time and memory do DPC and SSC introduce relative to the underlying GCN? Please distinguish this operator-level overhead from the useful accuracy–resource trade-offs obtained with separately selected architectures.

3 comments
AuthorsOpenReview ↗

1. Local degree-coordinate effects

Thank you for the careful reading and independent check of Proposition 3. We agree that geometric non-redundancy, predictive utility, and architecture-level resource trade-offs are different claims. The additional evidence separates them.

We provide run-level differences and confidence intervals for all seven datasets in Table 6, focusing our analysis on Computer, WikiCS and Minesweeper as requested. We further examine how the local correction affects nodes with different structural discrepancies.

Run-level performance and uncertainty. The table compares full DPC with Graph-wide and Shuffled-centered. Each cell reports the mean difference; [five run-level differences]; [marginal 95% CI], in percentage points. Differences are DPC-GCN minus the control, so positive values favor DPC-GCN.

Comparison: mean Δ; [five Δs]; [95% CI]ComputerWikiCSMinesweeperRatingsRoman-EmpireQuestionsTolokers
DPC-GCN − Graph-wide+0.189; [-0.1818, +0.3999, +0.1818, +0.3635, +0.1818]; [-0.097, +0.475]+0.195; [+0.1368, +0.7867, +0.1881, -0.1026, -0.0342]; [-0.242, +0.632]+0.415; [+0.2564, +0.8234, +0.2044, +0.5492, +0.2432]; [+0.085, +0.746]+0.134; [+0.0817, +0.3266, +0.5063, -0.3593, +0.1143]; [-0.269, +0.537]-0.011; [+0.6001, -0.2471, +0.1588, -0.2118, -0.3530]; [-0.498, +0.476]-0.020; [-0.0168, -0.0494, -0.0409, -0.0100, +0.0179]; [-0.053, +0.013]+0.038; [-0.0825, -0.1178, -0.1088, +0.3015, +0.1998]; [-0.207, +0.284]
DPC-GCN − Shuffled-centered+0.225; [+0.0727, +0.1091, +0.5453, +0.2181, +0.1818]; [-0.008, +0.459]+0.339; [+0.0513, +0.8893, +0.3421, +0.2052, +0.2052]; [-0.064, +0.742]+0.238; [+0.8949, +0.2147, -0.2672, -0.1647, +0.5114]; [-0.359, +0.835]-0.033; [+0.2450, -0.3266, +0.1960, -0.6369, +0.3593]; [-0.565, +0.499]-0.049; [+0.5648, +0.1765, -0.3883, -0.5471, -0.0529]; [-0.602, +0.503]+0.037; [-0.0125, +0.1028, -0.0584, +0.1009, +0.0523]; [-0.051, +0.125]+0.082; [-0.1684, -0.0856, +0.1750, +0.3086, +0.1812]; [-0.167, +0.331]

Runs are paired by their predefined run indices and corresponding data splits; 95% confidence intervals use paired t statistics on the run-level differences.

On the three datasets highlighted in your question, full DPC has higher mean performance than both controls. On Minesweeper, all five differences from Graph-wide are positive, with a mean gain of 0.415 points. On Computer and WikiCS, all five differences from Shuffled-centered are also positive, with mean gains of 0.225 and 0.339 points, respectively. The 95% paired interval excludes zero for Minesweeper versus Graph-wide; the remaining intervals include zero. These results show positive mean gains from the local coordinate on the highlighted tasks, with the clearest evidence on Minesweeper versus Graph-wide. Node-level effects. To further examine the contribution of the local correction, we hold all other model parameters fixed and remove it by setting λl=0\lambda_l=0 in every layer at inference. Test nodes are grouped by the structural discrepancy coordinate qiq_i, without using labels or predictions; larger qiq_i indicates a larger relative discrepancy between the node's normalized propagation rows. The table reports the mean increase in negative log-likelihood (NLL) after removal over five runs per dataset; positive values favor retaining the local term. Equal-qiq_i nodes remain together, yielding two groups on Minesweeper.

Retain local vs remove local: NLL reduction (nats)ComputerWikiCSMinesweeper
All test nodes-0.000098+0.002552+0.000149
Low-qq group-0.000484-0.001651-0.000037
Middle-qq group-0.000499-0.001175—
High-qq group+0.000671+0.010510+0.002335

Across all five runs on each of the three datasets, removing the local term while holding the other model parameters fixed increases mean NLL in the high-qq group. This finding provides node-level evidence of the role of local calibration, revealing prediction responses associated with structural discrepancy beyond aggregate performance metrics. It also motivates further investigation of the relationship between local topology and calibration benefits.

AuthorsOpenReview ↗

2. DPC beyond SSC

To distinguish DPC's contribution from that of the combined architecture, we compare four configurations with the same backbone: Backbone, DPC-GCN (Backbone + DPC), Backbone + SSC, and DPC-GCN-S (Backbone + DPC + SSC). Backbone uses standard GCN propagation. Normalization, residual connections, width, depth, dropout and training budget are unchanged; only DPC and SSC are enabled or disabled. Results are five-run means ± standard deviations on a percentage scale: ROC-AUC for Minesweeper, Questions and Tolokers, and accuracy otherwise.

Homophilous datasets

Model / controlCoraCiteSeerPubMedComputerPhotoCSPhysicsWikiCS
Backbone84.46 ± 0.6072.40 ± 0.4781.30 ± 0.5793.74 ± 0.1695.88 ± 0.2196.16 ± 0.1497.47 ± 0.1080.17 ± 0.57
DPC-GCN85.22 ± 0.9073.28 ± 0.6081.36 ± 0.5194.20 ± 0.1096.54 ± 0.1296.26 ± 0.1097.47 ± 0.0580.90 ± 0.28
Backbone + SSC84.98 ± 0.4472.76 ± 0.6281.58 ± 0.3393.86 ± 0.1995.59 ± 0.2196.15 ± 0.1097.43 ± 0.0480.29 ± 0.23
DPC-GCN-S85.26 ± 0.6473.08 ± 0.5481.42 ± 0.3994.21 ± 0.2296.51 ± 0.1396.25 ± 0.1197.46 ± 0.0580.83 ± 0.25

Heterophilous datasets

Model / controlRatingsMinesweeperRoman-EmpireQuestionsTolokers
Backbone54.21 ± 0.5197.60 ± 0.4991.38 ± 0.3579.06 ± 0.4986.37 ± 0.54
DPC-GCN55.42 ± 0.3897.73 ± 0.3891.58 ± 0.2279.06 ± 0.5486.41 ± 0.56
Backbone + SSC54.35 ± 0.1698.58 ± 0.3291.97 ± 0.1679.89 ± 0.6386.52 ± 0.57
DPC-GCN-S55.55 ± 0.3998.73 ± 0.1991.97 ± 0.2279.89 ± 0.6386.62 ± 0.53

DPC without SSC. DPC-GCN achieves higher means than Backbone on 11 of 13 datasets, with essentially unchanged performance on Physics and Questions. Mean gains on Cora, CiteSeer, Computer, Photo, WikiCS and Amazon-Ratings are approximately 0.76, 0.88, 0.46, 0.66, 0.73 and 1.21 percentage points, respectively. This comparison directly demonstrates degree-calibration gains without SSC.

DPC with SSC retained. Comparing DPC-GCN-S with Backbone + SSC, DPC still adds approximately 0.92, 0.54, 0.35 and 1.20 points on Photo, WikiCS, Computer and Amazon-Ratings, respectively. As you noted, the additional gains on Roman-Empire and Questions are essentially zero, and those on Minesweeper and Tolokers are small. Thus, DPC's benefit with SSC already present is dataset-dependent, with the four datasets above providing more direct positive examples.

SSC and the combined model. Comparing DPC-GCN-S with DPC-GCN, SSC improves mean performance on all five heterophilous datasets, including gains of approximately 1.00 and 0.83 points on Minesweeper and Questions; changes on homophilous datasets are only −0.20 to +0.06 points. These results distinguish the components' contributions: the backbone comparison supports DPC on its own, the SSC-retained comparison shows DPC's additional value on some tasks, and the larger gains on several heterophilous datasets primarily reflect SSC. The combined model's overall performance is interpreted through these component comparisons rather than attributed entirely to DPC.

3. Model-selection effort

Thank you for highlighting the importance of model-selection effort. Our tuning proceeds in stages with coordinate-wise refinement rather than a full Cartesian grid. For baselines, we prioritize released configurations that match the current evaluation setup and conduct supplementary searches when applicable settings are unavailable.

Taking Cora as an example, the corresponding DPC-GCN search stages consider hidden widths H∈{512,1024,2048}H\in\{512,1024,2048\}, depths L∈{2,3,4}L\in\{2,3,4\}, dropout rates D∈{0.3,0.4,0.5}D\in\{0.3,0.4,0.5\}, and coefficient initializations α0∈{0,0.1,…,1}\alpha_0\in\{0,0.1,\ldots,1\} and λ0∈{−0.5,−0.25,0,0.25,0.5}\lambda_0\in\{-0.5,-0.25,0,0.25,0.5\}. Backbone parameters are varied one axis at a time around (1024,3,0.4)(1024,3,0.4), followed by separate initialization searches. DPC-GCN-S additionally considers SSC learning-rate multipliers mτ∈{0.1,0.25,0.5,1,2,4}m_\tau\in\{0.1,0.25,0.5,1,2,4\} at a fixed DPC configuration; layer-wise calibration coefficients are learned during training.

For comparison, Tuned-GCN follows the Classic GNN paper's Cora configuration: hidden width 512, three layers, dropout 0.7, learning rate 0.001, and 500 epochs, without normalization or residual connections. This setting incorporates the original study's tuning over hidden widths {64,256,512}\{64,256,512\}, depths {1,…,10}\{1,\ldots,10\}, dropout rates {0.2,0.3,0.5,0.7}\{0.2,0.3,0.5,0.7\}, learning rates {0.001,0.005,0.01}\{0.001,0.005,0.01\}, and choices of normalization and residual connections. Reusing a released configuration therefore differs from using untuned defaults.

For baselines requiring adaptation to the current split or evaluation setup, we start from released configurations and perform supplementary searches. CPF on Cora retains its two-layer feature network and refines initial configurations coordinate-wise over hidden widths {64,128,256}\{64,128,256\}, filter orders {5,10,20}\{5,10,20\}, feature dropout {0,0.3,0.5,0.7}\{0,0.3,0.5,0.7\}, and filter dropout {0,0.3,0.5}\{0,0.3,0.5\}. Feature-network learning rates are {0.001,0.005,0.01,0.05}\{0.001,0.005,0.01,0.05\}, with 0.10.1 additionally considered for the filtering branches. Weight decay covers {0,0.0005,0.005}\{0,0.0005,0.005\} for the feature network, {0,0.0005}\{0,0.0005\} for the structural-filter branch, and {0,0.0005,0.05}\{0,0.0005,0.05\} for the feature-filter branch.

H2GCN retains the public configuration's hidden width of 64 and two propagation rounds, and considers dropout {0,0.5}\{0,0.5\}, learning rates {0.001,0.005,0.01}\{0.001,0.005,0.01\}, weight decay {0,0.00005,0.0005}\{0,0.00005,0.0005\}, and ReLU/identity activations. These ranges describe the values considered in the respective search stages, not an exhaustive evaluation of all combinations.

The inherited tuning effort and local search scopes do differ across models. Appendices C.3–C.4 document the configuration sources, settings, and search ranges. We will further refine hyperparameter tuning during the rebuttal period by expanding candidates over relevant backbone and method-specific hyperparameters for baselines with limited coverage in existing configurations and supplementary searches. Selection will use validation performance to better assess their performance potential and the sensitivity of the current comparisons to the search scope.

4. Operator overhead at matched architecture

Using the five heterophilous datasets as examples, we measure the additional training time and GPU memory introduced by DPC and SSC under matched backbone configurations. Within each dataset, Backbone, DPC-GCN, and DPC-GCN-S share the same hidden width, depth, dropout, and training protocol, differing only in the calibration modules. Measurements use an exclusive RTX 5090. Each cell reports training time (ms/epoch) / peak allocated GPU memory (MiB), taking the median across three measurement blocks with rotated model order, each comprising 10 warmup and 50 measured training steps. Timing excludes graph preprocessing; memory includes the data, model, optimizer, cached graph, and training activations.

ModelRatingsMinesweeperRoman-EmpireQuestionsTolokers
Backbone7.37 / 731.810.36 / 173.611.25 / 962.369.70 / 4450.74.48 / 200.8
DPC-GCN8.49 / 739.119.75 / 189.513.44 / 974.076.71 / 4511.58.09 / 274.5
DPC-GCN-S11.99 / 1029.724.28 / 260.616.95 / 1373.4114.01 / 7074.610.28 / 329.2

Across these five datasets, DPC-GCN takes 1.10–1.91× the backbone's training time; adding SSC takes a further 1.23–1.49× relative to DPC-GCN. The first two rows isolate the overhead of DPC, and the last two isolate the additional overhead of SSC. These matched-architecture comparisons directly quantify the additional cost of calibration. Section 4.5 instead compares accuracy and resource use at each method's selected configuration, rather than claiming an operator-level speedup over the same GCN backbone; we will make this distinction explicit in the revision.

Reviewer_E9FNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DPC-GCN, which lets each layer move between GCN's symmetric normalization and mean aggregation, using a shared coefficient and an additional receiver-dependent correction based on a fixed degree statistic. An optional step (SSC) introduces a single learnable signed scalar per layer, allowing the model to either smooth or sharpen the aggregated features. The authors provide supporting theory and evaluate the method on 16 node classification benchmarks, reporting results that match or exceed the best baseline mean on all eight homophilous datasets and lead on four of five heterophilous datasets.

Critical strengths

The method is simple and computationally efficient, adding only two or three trainable scalars per layer while retaining sparse graph propagation. I particularly liked the SSC component. The learned signs make sense across different datasets: sharpening on Minesweeper, Roman-Empire, and Questions, and smoothing on Tolokers. SSC also brings noticeable gains on heterophilous datasets in the ablations, around one percentage point on Minesweeper and Questions.

The experimental evaluation is fairly thorough and well documented, with matched ablations, design controls, and detailed configurations. These make it easier to understand where the improvements come from and to reproduce the results.

Critical weaknesses
  1. Limited performance gains.

The improvements over the strongest baselines are often small relative to the reported standard deviations, without significance tests to establish their reliability. The ablations also show no additional gain from DPC on Roman-Empire and Questions at the reported precision, while SSC contributes little on homophilous graphs. A matched plain-backbone baseline with both DPC and SSC removed would help clarify their actual contributions.

  1. Limited novelty and missing comparisons.

DPC is closely related to existing degree normalization methods, such as learning the exponent in D~−rA~D~r−1\tilde D^{-r}\tilde A\tilde D^{r-1}, while SSC resembles a learnable first-order graph filter. Although AdaGNN-S and CPF are included, comparisons with simpler alternatives and methods such as GPR-GNN and FAGCN are missing. This makes it difficult to judge how much the proposed mechanisms add beyond existing approaches.

Questions for authors

Could the authors report significance tests or confidence intervals for the main comparisons, and include a plain-backbone baseline with both DPC and SSC removed under the same architecture and training setup?

How does DPC compare with simply learning the normalization exponent rr? For SSC, how does it compare with existing adaptive filtering methods such as GPR-GNN and FAGCN?

Since the receiver-dependent coefficient can fall outside [0,1][0,1], some calibrated edge weights could potentially become negative. How often does this happen, and could it contribute to the gains on heterophilous graphs?

3 comments
AuthorsOpenReview ↗

1. Statistical reliability and matched-backbone controls

Thank you for the constructive feedback.

We compare DPC-GCN-S with six strong baselines, including previously reported SOTA methods. Cells report Δ (the difference between the two-decimal means in the paper, percentage points) and unadjusted two-sided Welch t-test p-values from five runs per method; * marks Holm-adjusted p<0.05. Tests use original run scores or archived summaries with sample variances; those involving rounded data are approximate.

W/T/L counts wins/ties/losses using two-decimal means across all available datasets. Overall p is from a two-sided exact sign test, excluding ties.

DPC-GCN-S — Homophilous datasets

BaselineCoraCiteSeerPubMedComputerPhotoCSPhysicsWikiCS
Tuned-GCN+0.72; p=0.1887+0.44; p=0.2178+0.72; p=0.1810+0.35; p=0.0312+0.58; p=0.0101+0.61; p=0.0018-0.01; p=0.8295+0.80; p=0.0275
Tuned-GAT+2.86; p=0.0012+1.22; p=0.0070+1.66; p=0.0342+0.48; p=0.0087+0.01; p=0.9547+0.09; p=0.1988+0.33; p=0.0072+0.17; p=0.6523
Polynormer+9.84; p=2.42e-08*+8.34; p=6.60e-06*+6.00; p=9.71e-06*+0.50; p=0.0336+0.01; p=0.9482+0.79; p=0.0003*+0.35; p=3.16e-05*+1.00; p=0.0746
GT-SNT+0.50; p=0.2756+1.90; p=0.0380+1.48; p=0.0993+1.37; p=3.95e-05*+1.35; p=5.28e-06*+1.36; p=5.53e-07*+0.23; p=0.0002*+1.35; p=0.0207
H2GCN+1.38; p=0.0113+1.62; p=0.0021+0.72; p=0.0424+3.01; p=4.52e-08*+0.63; p=0.0003*+0.32; p=0.0047+0.33; p=7.26e-05*+1.66; p=0.0038

DPC-GCN-S — Heterophilous datasets

BaselineRatingsMinesweeperRoman-EmpireQuestionsTolokersAll-data W/T/LOverall p (Holm-adjusted)
Tuned-GCN+1.80; p=0.0002*+1.46; p=0.0001*+0.84; p=0.0005*+1.27; p=0.0194+0.61; p=0.128512/0/10.0103
Tuned-GAT+0.24; p=0.5147+1.47; p=0.0174+1.90; p=0.0006*+2.35; p=0.0008*+0.84; p=0.031413/0/00.0017
Polynormer+1.09; p=0.0026+1.64; p=3.01e-05*-0.54; p=0.0190+1.74; p=0.0521OOM11/0/10.0127
GT-SNT+6.25; p=1.24e-08*+10.90; p=2.27e-07*+16.71; p=2.40e-08*+2.63; p=0.0036+4.60; p=1.49e-05*13/0/00.0017
H2GCN+8.32; p=2.30e-09*+8.00; p=3.56e-10*+13.38; p=5.31e-09*+5.02; p=0.0004*+3.94; p=3.50e-05*13/0/00.0017
Co-GNN+1.71; p=0.0007*+2.80; p=1.49e-05*+1.87; p=5.65e-05*+1.25; p=0.1539+2.62; p=0.00275/0/00.0625

DPC-GCN-S has higher means on most datasets against each baseline. Five of six overall comparisons remain significant after Holm correction (p=0.0017–0.0127), supporting a consistent cross-dataset advantage. Dataset-level significance is marked in the tables.

Matched-backbone controls.

The same-architecture control (Backbone) retains our normalization, residual paths, width, depth, dropout and training budget, with DPC and SSC disabled. Scores are five-run mean ± SD (%, divisor 5), consistent with the corresponding paper rows. DPC-GCN, Backbone + SSC, and DPC-GCN-S reuse the paper results.

Homophilous datasets

Model / controlCoraCiteSeerPubMedComputerPhotoCSPhysicsWikiCS
Backbone84.46 ± 0.6072.40 ± 0.4781.30 ± 0.5793.74 ± 0.1695.88 ± 0.2196.16 ± 0.1497.47 ± 0.1080.17 ± 0.57
DPC-GCN85.22 ± 0.9073.28 ± 0.6081.36 ± 0.5194.20 ± 0.1096.54 ± 0.1296.26 ± 0.1097.47 ± 0.0580.90 ± 0.28
Backbone + SSC84.98 ± 0.4472.76 ± 0.6281.58 ± 0.3393.86 ± 0.1995.59 ± 0.2196.15 ± 0.1097.43 ± 0.0480.29 ± 0.23
DPC-GCN-S85.26 ± 0.6473.08 ± 0.5481.42 ± 0.3994.21 ± 0.2296.51 ± 0.1396.25 ± 0.1197.46 ± 0.0580.83 ± 0.25

Heterophilous datasets

Model / controlRatingsMinesweeperRoman-EmpireQuestionsTolokers
Backbone54.21 ± 0.5197.60 ± 0.4991.38 ± 0.3579.06 ± 0.4986.37 ± 0.54
DPC-GCN55.42 ± 0.3897.73 ± 0.3891.58 ± 0.2279.06 ± 0.5486.41 ± 0.56
Backbone + SSC54.35 ± 0.1698.58 ± 0.3291.97 ± 0.1679.89 ± 0.6386.52 ± 0.57
DPC-GCN-S55.55 ± 0.3998.73 ± 0.1991.97 ± 0.2279.89 ± 0.6386.62 ± 0.53

DPC-GCN has higher means than Backbone on 11/13 datasets; Physics and Questions tie at the displayed precision. Adding SSC improves all five heterophilous means by 0.13–1.00 points, versus −0.20 to +0.06 on homophilous graphs. DPC thus benefits most datasets, while SSC mainly benefits heterophilous graphs.

AuthorsOpenReview ↗

2. Learnable normalization and adaptive filters

DPC versus a learned degree-normalization exponent. Both use source–receiver degrees: exponent normalization applies a shared rule, while DPC combines shared and neighborhood-dependent corrections. Let sij=ϕj−ϕis_{ij}=\phi_j-\phi_i, where ϕi=12log⁡d~i\phi_i=\tfrac12\log\widetilde d_i. Relative to symmetric GCN, the two edge factors are

\frac{(P_{\mathrm{DPC}}^{(l)})_{ij}}{(P_g)_{ij}}=1+(\alpha_l+\lambda_l t_i)(\exp(s_{ij})-1),$$ on existing edges, with $P_r=\widetilde D^{-r_l}\widetilde A\widetilde D^{r_l-1}$. DPC differs from a shared exponent in three respects. (i) Exponent normalization uses $r_l$ to change the exponential response to degree-potential differences uniformly; DPC instead learns the shared strength $\alpha_l$ and local-offset scale $\lambda_l$ along the degree-potential-induced correction $\exp(s_{ij})-1$. (ii) A shared exponent assigns the same relative calibration factor to the same degree ratio. DPC's $t_i$ summarizes degree-potential discrepancies across the receiver's entire neighborhood, so edges with the same non-unit degree ratio can receive different calibration according to their neighborhood profiles. (iii) With $\Delta=\widetilde D^{-1}\widetilde A-P_g$, discrepancy-weighted centering gives $\langle\Delta,\operatorname{Diag}(t)\Delta\rangle_F=0$, so a nonzero local correction cannot be replaced by changing the shared mixing coefficient (Proposition 1). DPC can therefore adapt to neighborhood differences while learning a graph-wide degree-weighting trend, without per-node trainable parameters: it uses only two scalars per layer and retains sparse propagation on the original graph. We replace DPC with a per-layer learned exponent in the same backbone and disable SSC, making DPC-GCN the direct comparator. Exponent free directly tests an unconstrained learned normalization exponent; Exponent [0,1] is an additional bounded control whose range includes column normalization ($r=0$), symmetric GCN ($r=0.5$), and row-mean propagation ($r=1$). For both, we select the initialization from $r_0\in\{0.5,1\}$ by validation performance. All scores below are five-run means ± sample SD. **Homophilous datasets: DPC versus exponent normalization** | Model / control | Cora | CiteSeer | PubMed | Computer | Photo | CS | Physics | WikiCS | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | DPC-GCN | 85.22 ± 0.90 | 73.28 ± 0.60 | 81.36 ± 0.51 | 94.20 ± 0.10 | 96.54 ± 0.12 | 96.26 ± 0.10 | 97.47 ± 0.05 | 80.90 ± 0.28 | | Exponent [0,1] | 84.66 ± 0.67 | 72.60 ± 0.66 | 79.76 ± 0.79 | 94.07 ± 0.11 | 96.38 ± 0.19 | 96.09 ± 0.10 | 97.43 ± 0.05 | 80.57 ± 0.42 | | Exponent free | 84.88 ± 0.57 | 72.60 ± 0.66 | 79.82 ± 0.66 | 93.94 ± 0.22 | 95.91 ± 0.29 | 96.09 ± 0.10 | 97.43 ± 0.05 | 80.59 ± 0.45 | DPC-GCN has a higher mean than both exponent controls on all eight homophilous datasets. **Adaptive filters.** On the five heterophilous datasets, we compare complete GPR-GNN and FAGCN models, with validation-selected configurations, and replace only SSC with their core filters on the unchanged DPC backbone. DPC+GPR uses $K=10$ with PPR initialization (restart probability 0.1); DPC+FAG uses one update $0.3C+A_{\mathrm{gate}}(C)C$. Feature transformations, residual paths, normalization, backbone dropout, classifier, width/depth and epoch budgets are retained, with no extra feature network or dropout. Complete GPR-GNN and FAGCN use sparse implementations of their released architectures and the same splits, metrics and epoch budgets; for each method and dataset, the additional tuning round selects among 25 candidate configurations by validation performance before five-run evaluation. **Heterophilous datasets: SSC versus adaptive filters** | Model / control | Ratings | Minesweeper | Roman-Empire | Questions | Tolokers | | --- | --- | --- | --- | --- | --- | | DPC-GCN-S | 55.55 ± 0.39 | 98.73 ± 0.19 | 91.97 ± 0.22 | 79.89 ± 0.63 | 86.62 ± 0.53 | | GPR-GNN | 51.28 ± 0.50 | 89.70 ± 0.32 | 73.42 ± 0.96 | 73.99 ± 1.37 | 76.40 ± 1.47 | | FAGCN | 51.66 ± 0.51 | 90.05 ± 0.34 | 74.70 ± 0.79 | 77.35 ± 1.01 | 81.63 ± 0.43 | | DPC + GPR | 48.34 ± 0.47 | 53.20 ± 2.31 | 92.14 ± 0.30 | 66.12 ± 1.19 | 84.28 ± 1.84 | | DPC + FAG | 48.25 ± 0.74 | 63.93 ± 6.78 | 63.97 ± 3.50 | 66.89 ± 1.34 | 83.27 ± 0.32 | Among the tested core-filter replacements, SSC has the highest mean on four of five datasets, while GPR is higher on Roman-Empire by 0.17 points. SSC uses one scalar and one additional propagation per layer, versus 11 coefficients and ten propagations for GPR, or a feature-dependent edge gate for FAG. These comparisons support SSC's effectiveness in the tested backbone and configurations.
AuthorsOpenReview ↗

3. Negative weights

We measured the actual propagation weights of DPC-GCN on 13 datasets to separate DPC from SSC. At the validation-selected checkpoints from five runs per dataset, eight datasets have no negative edges in any layer. Among directed non-self propagation edges, the maximum negative-edge count and absolute-weight fractions across datasets, runs and layers are approximately 0.016535% and 0.002322%, respectively. The latter is the sum of absolute negative weights divided by the sum of all absolute weights. Thus, negative propagation is rare in these selected models.

We additionally train DPC-GCN with wij←max⁡(wij,0)w_{ij}\leftarrow\max(w_{ij},0) throughout training and inference, without renormalizing the weights. SSC is disabled in both models, with backbone architecture, hyperparameters, splits and epoch budgets unchanged. Compared with the original model results in the paper, the absolute difference in five-run mean test performance is below 0.1 percentage points on 11/13 datasets; the remaining two, WikiCS and Ratings, decrease by 0.212 and 0.173 points, respectively. Overall, these metric changes are small.

Given their rarity, we further examine whether negative messages nevertheless have local effects. On the five datasets with negative edges, we use five runs under the same configurations and, at each fixed checkpoint, set negative weights to zero only at inference while keeping all model parameters and other weights unchanged, without renormalization. Direct receivers are identified from the original negative edges, not selected by prediction outcomes. Retaining negative weights increases true-class probabilities and probability margins for some nodes, with positive mean changes over receivers on four datasets. However, the magnitude and direction vary across nodes, and no receiver changes its final predicted class under this intervention.

Local intervention on negative-edge receivers

Model / statisticComputerPhysicsWikiCSRoman-EmpireTolokers
DPC-GCN: Δ true-class probability (pp)+0.000070+0.014274+0.069488+0.011602−0.002051
DPC-GCN: Δ probability margin (pp)+0.000139+0.030283+0.130358+0.030061−0.004101
Unique receiver nodes56631104
Runs with receivers / all runs2/55/51/55/53/5

Here, differences are retained minus zeroed negative weights, in percentage points, averaged first within each nonempty run and then across runs. The probability margin is the true-class probability minus the largest incorrect-class probability. All direct receivers are included across training, validation and test splits for this local mechanism analysis; node counts are deduplicated across runs.

These results show that negative messages can make positive local contributions, but do not establish them as the main source of the overall performance gains. Among the five heterophilous datasets, only Roman-Empire and Tolokers exhibit negative edges in the examined DPC-GCN checkpoints. Their role across different graph structures and throughout training merits further study.

Comments on the paper

5 comments

Thank you for the constructive feedback.

1. Statistical uncertainty

We compare DPC-GCN-S with six strong baselines, including previously reported SOTA methods. Cells report Δ (the difference between the two-decimal means in the paper, percentage points) and unadjusted two-sided Welch t-test p-values from five runs per method; * marks Holm-adjusted p<0.05. Tests use original run scores or archived summaries with sample variances; those involving rounded data are approximate.

W/T/L counts wins/ties/losses using two-decimal means across all available datasets. Overall p is from a two-sided exact sign test, excluding ties.

DPC-GCN-S — Homophilous datasets

BaselineCoraCiteSeerPubMedComputerPhotoCSPhysicsWikiCS
Tuned-GCN+0.72; p=0.1887+0.44; p=0.2178+0.72; p=0.1810+0.35; p=0.0312+0.58; p=0.0101+0.61; p=0.0018-0.01; p=0.8295+0.80; p=0.0275
Tuned-GAT+2.86; p=0.0012+1.22; p=0.0070+1.66; p=0.0342+0.48; p=0.0087+0.01; p=0.9547+0.09; p=0.1988+0.33; p=0.0072+0.17; p=0.6523
Polynormer+9.84; p=2.42e-08*+8.34; p=6.60e-06*+6.00; p=9.71e-06*+0.50; p=0.0336+0.01; p=0.9482+0.79; p=0.0003*+0.35; p=3.16e-05*+1.00; p=0.0746
GT-SNT+0.50; p=0.2756+1.90; p=0.0380+1.48; p=0.0993+1.37; p=3.95e-05*+1.35; p=5.28e-06*+1.36; p=5.53e-07*+0.23; p=0.0002*+1.35; p=0.0207
H2GCN+1.38; p=0.0113+1.62; p=0.0021+0.72; p=0.0424+3.01; p=4.52e-08*+0.63; p=0.0003*+0.32; p=0.0047+0.33; p=7.26e-05*+1.66; p=0.0038

DPC-GCN-S — Heterophilous datasets

BaselineRatingsMinesweeperRoman-EmpireQuestionsTolokersAll-data W/T/LOverall p (Holm-adjusted)
Tuned-GCN+1.80; p=0.0002*+1.46; p=0.0001*+0.84; p=0.0005*+1.27; p=0.0194+0.61; p=0.128512/0/10.0103
Tuned-GAT+0.24; p=0.5147+1.47; p=0.0174+1.90; p=0.0006*+2.35; p=0.0008*+0.84; p=0.031413/0/00.0017
Polynormer+1.09; p=0.0026+1.64; p=3.01e-05*-0.54; p=0.0190+1.74; p=0.0521OOM11/0/10.0127
GT-SNT+6.25; p=1.24e-08*+10.90; p=2.27e-07*+16.71; p=2.40e-08*+2.63; p=0.0036+4.60; p=1.49e-05*13/0/00.0017
H2GCN+8.32; p=2.30e-09*+8.00; p=3.56e-10*+13.38; p=5.31e-09*+5.02; p=0.0004*+3.94; p=3.50e-05*13/0/00.0017
Co-GNN+1.71; p=0.0007*+2.80; p=1.49e-05*+1.87; p=5.65e-05*+1.25; p=0.1539+2.62; p=0.00275/0/00.0625

DPC-GCN-S has higher means on most datasets against each baseline. Five of six overall comparisons remain significant after Holm correction (p=0.0017–0.0127), supporting a consistent cross-dataset advantage. Dataset-level significance is marked in the tables.

2. Generality of Figure 1(a)

Thank you for suggesting this extension. Figure 1(a) asks whether neighborhoods with matched receiver degrees differ in their sensitivity to degree-based message weighting. We repeated the diagnostic on all 13 medium-scale datasets and ogbn-arxiv, keeping the original feature normalization, relative GCN–mean message-change measure, and 12-bin matching procedure. Each retained degree–bin cell contains at least five nodes, and weighting makes receiver-degree distributions identical across bins. No training or labels are used.

The tables report the weighted median relative message change in the lowest → highest degree-contrast bins. Message changes are dimensionless; only retained-node coverage is expressed as a percentage.

Homophilous datasets

Diagnostic statisticCoraCiteSeerPubMedComputerPhotoCSPhysicsWikiCS
Retained nodes (%)74.5653.5097.5580.2079.4889.9594.3467.66
Low → high contrast: median message change0.035 → 0.2580.019 → 0.2300.098 → 0.3340.052 → 0.4890.053 → 0.4730.054 → 0.2480.048 → 0.2510.101 → 0.504

Heterophilous and large-scale datasets

Diagnostic statisticRatingsMinesweeperRoman-EmpireQuestionsTolokersogbn-arxiv
Retained nodes (%)97.970.0057.5496.4759.4886.11
Low → high contrast: median message change0.025 → 0.307N/A0.029 → 0.0740.109 → 0.3920.223 → 0.5930.039 → 0.454

All 13 graphs with sufficient matching support show larger median message changes in the highest-contrast bin than in the lowest, extending the PubMed observation to heterophilous and large-scale graphs. Minesweeper's near-regular grid lacks sufficient within-degree contrast variation for this procedure; its median contrast is zero and its median message change is approximately 3.4×10−83.4\times10^{-8}. N/A therefore indicates insufficient matching support, not a missing run.

We also repeated the analysis with five node-feature permutations and five independent Gaussian-feature draws per graph, retaining the same normalization, bins, and matching weights. Averaged over these repetitions, both controls preserve the low-to-high contrast relationship on all 13 supported graphs. This indicates a topology-induced component rather than a PubMed-specific feature effect. The diagnostic establishes sensitivity of message construction; whether adapting these messages benefits classification is addressed by the predictive analyses below.

We will add these cross-dataset diagnostics, random-feature controls, and corresponding visualizations to the revised paper to clarify the scope of the phenomenon in Figure 1(a).

3. From degree-potential corrections to classification

Thank you for this question. We examined the relationship between structural discrepancies and DPC's classification gains across all 13 datasets, and further analyzed message corrections within the trained models. Gains occur on graphs with both smaller and larger structural discrepancies, rather than following a simple monotonic relationship. Below, we use Computer and WikiCS as two illustrative cases: they learn opposite receiver-dependent modulation patterns, allowing us to examine how degree-potential corrections affect classification through different message-passing behaviors.

For each of these two cases, we used five validation-selected DPC-GCN models trained under the paper configurations, with SSC disabled. The table reports inference interventions within fixed models. “Total” compares the full model with both corrections disabled; “local” retains the shared correction and disables only the local term. All model parameters remain fixed.

Computer and WikiCS learn opposite group-level adjustments. Low/high qq are the bottom/top graph-defined tertile groups of the topology-discrepancy coordinate, with ties retained. Coefficient ranges span layer-wise five-run group means, not node extrema or confidence intervals.

Modulation / fixed-model effectComputerWikiCS
Low-qq mean local coefficient λlti\lambda_l t_i[−0.175, −0.124][+0.272, +0.278]
High-qq mean local coefficient λlti\lambda_l t_i[+0.023, +0.032][−0.052, −0.051]
Fixed model: total accuracy effect (pp)+1.3668+0.8278
Fixed model: additional local accuracy effect (pp)+0.1018+0.1437

On average, Computer's local term opposes the shared correction at low-qq positions and reinforces it at high-qq positions; WikiCS learns the reverse pattern. Total accuracy improves in all five runs on both datasets. The local effect is positive in three runs and zero in two on Computer, and positive in all five on WikiCS. These examples illustrate task-dependent modulation rather than a fixed rule to strengthen high-qq receivers.

To connect this modulation to classification, we integrated edge-message contributions through the original downstream network while increasing corrections from zero to their learned values. Their sum recovers the measured test negative log-likelihood (NLL) change within numerical tolerance. On WikiCS, with the shared correction retained, messages from higher-degree neighbors to high-qq receivers where the local term attenuates the shared correction contribute +0.002913 nats to NLL reduction, positive in all five runs. After including adverse contributions elsewhere, the net reduction is 0.002552 nats, alongside the +0.1437-point accuracy effect above; both improve in every run. These contributions concern downstream test predictions, not necessarily the receivers' own losses. This identifies a concrete route through which receiver-specific attenuation supports classification.

These two cases illustrate how DPC adapts information propagation to the task: degree-potential differences produce neighborhood-dependent message corrections, the local coordinate modulates their strength across receivers, and specific message paths connect these adjustments to classification responses. The complete 13-graph diagnostics also include negative and near-zero responses, indicating dataset-dependent effects. We will add the results for all datasets and the detailed attribution analysis to the revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.