acceptodds
Under review as a conference paper at ICLR 2027

Propagate, Then Sharpen: Post-Hoc Refinement of Frozen Node Classifiers

Abstract

We study post-hoc refinement of frozen node classifiers: given only the graph and class distributions predicted by a frozen model, can we improve accuracy without access to node features, model parameters, or gradients? APPNP answers this by propagating logits with a restart towards the initial predictions, minimizing the anchored Dirichlet energy. Instead, we consider the Potts energy, and decompose it into a Dirichlet term, which penalizes disagreement between neighbouring nodes, and a Gini term, which penalizes indecision within each node. This decomposition motivates Propagate, Then Sharpen (PtS), which alternates between propagation of class probabilities and node-wise, mass-preserving sharpening, with only one additional hyperparameter selected using labelled validation nodes. Across nine homophilic graphs, with a frozen MLP backbone, PtS improves mean test accuracy over independently tuned APPNP by percentage points on clean inputs and under severe Gaussian feature corruption. Gains over APPNP become smaller, but remain positive with frozen GCN and GraphSAGE backbones. Sharpening also removes most of the accuracy loss of deep propagation: on clean inputs without restart, accuracy falls by points between and propagation steps under PtS, compared with for APPNP.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph neural networks, graph propagation, post-hoc refinement, node classification, Potts, energy-based methods

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

This paper studies post-hoc refinement of frozen node classifiers using only the graph G and predicted class distributions Q. The method requires no node features, model parameters, or training gradients. The authors motivate propagation and node wise sharpening through a Dirichlet–Gini decomposition of the Potts interaction. Their method, PtS, first propagates probabilities with restart and then restores its row mass. Setting the additional parameter η to zero recovers PPR-Prob exactly. The authors prove that sharpening preserves the current class ranking and decreases a single-node KL–Gini energy.

Critical strengths
  1. The method is simple, has clear input requirements, and provides a useful theoretical explanation. PtS uses only G and Q and adds one parameter that can be disabled, making it applicable when model updates are unavailable. Restoring row mass recovers the original probability propagation at η = 0. The local majorization proof explains sharpening’s effect, and the two-community construction illustrates its interaction with propagation. The authors distinguish local energy descent from full-graph descent, keeping the theoretical claims within their demonstrated scope.
Critical weaknesses
  1. The specific sharpening formula lacks direct controls.The paper does not compare temperature sharpening (normalized p^γ) or hard thresholding under identical propagation, row-mass handling, and tuning budgets. Given prior diffusion–thresholding and Potts methods, these controls are needed to establish whether the gains depend on the KL–Gini design or on sharpening generally.

  2. Sharpening’s independent gains are modest and dataset-dependent. Gains over independently tuned PPR-Prob were 0.57 points on clean data and 1.09 under severe noise. Reaction OFF’s 5.31-point loss uses PtS-selected propagation parameters and does not establish the independent gain. Paired uncertainty estimates are needed to distinguish sharpening’s contribution from probability-space propagation.

  3. Global sharpening can reinforce incorrect predictions. Table 18 reports negative gains in two low-homophily bins, and PtS underperformed APPNP on severely noisy products. Sharpening preserves rankings, while Proposition 2 starts from correct predictions. The prevalence of affected nodes and whether validation can flag these failures remain unclear, limiting recommendations for general use.

Questions for authors
  1. Under identical probability propagation, row-mass handling, and independent tuning budgets, how do temperature sharpening rγ(p)∝pγr\gamma(p) \propto p^\gamma and hard thresholding perform? Does the KL–Gini reaction retain a consistent advantage at both large depths and the depths selected in practice?

  2. If output temperature is fitted only on clean validation data and then tested at σ = 2, how do PtS and APPNP compare in NLL and ECE? Could you report per-graph results, including the three graphs omitted from Table 20, to assess whether average improvements mask calibration losses?

  3. What fraction of each graph’s test nodes falls into Table 18’s low-homophily bins, and are negative gains concentrated in particular graphs or types of incorrect priors? For severely noisy products, is the degradation mainly due to propagation space, sharpening, or selection bias? Can validation results flag this failure?

2 comments
AuthorsOpenReview ↗

Response Bxxx- Part 1/2

Thank you for the careful review and the precise questions, and for noting that our theoretical claims stay within their demonstrated scope. New material is marked in red in the revised PDF.

The specific sharpening formula lacks direct controls. The paper does not compare temperature sharpening (normalized p^γ\gamma) or hard thresholding under identical propagation, row-mass handling, and tuning budgets. Given prior diffusion–thresholding and Potts methods, these controls are needed to establish whether the gains depend on the KL–Gini design or on sharpening generally.

Under identical probability propagation, row-mass handling, and independent tuning budgets, how do temperature sharpening and hard thresholding perform?

The gains come from soft sharpening generally, not from the KL–Gini design specifically. With the same probability propagation, row-mass handling, off option and 250-trial validation budget, PtS does not significantly outperform temperature sharpening: the mean differences are +0.04+0.04 percentage points (95% CI: [−0.003,0.09][-0.003, 0.09]) on clean inputs and +0.02+0.02 points (95% CI: [−0.02,0.06][-0.02, 0.06]) at σ=2\sigma=2. Both paired bootstrap intervals include zero. Hard thresholding is 0.530.53 points (95% CI: [0.45,0.60][0.45, 0.60]) below PtS on clean inputs and 0.920.92 points (95% CI: [0.79,1.13][0.79, 1.13]) below it at σ=2\sigma=2 (Section 4.2, Table 2). The two soft rules are the Gini and the Shannon-entropy versions of one KL-regularized sharpening step, and hard thresholding is their common limit (new remark in Appendix B, and our answer to Reviewer s4Ta).

Does the KL–Gini reaction retain a consistent advantage at both large depths and the depths selected in practice?

At the selected depths, PtS does not show a statistically significant mean-accuracy advantage over temperature sharpening: the paired intervals for the differences of 0.040.04 and 0.020.02 percentage points both include zero. At K=100K=100 without restart, PtS, temperature sharpening and hard thresholding reach 73.15%, 73.14% and 73.76% on clean inputs, and 57.76%, 57.74% and 57.30% at σ=2\sigma=2 (Appendix O.1, Table 22). Against hard thresholding the advantage therefore holds at the selected depths, by 0.53 and 0.92 points, but not consistently at K=100K=100. Against propagation without sharpening the advantage holds at both. At the selected depths it is 0.57 and 1.09 points over PPR-Prob and 1.71 and 3.90 points over APPNP (Tables 1 and 2). At K=100K=100 without restart, PPR-Prob reaches 44.6% and 37.2%, and APPNP 40.5% and 34.7% (Figure 2, Table 21).

Sharpening’s independent gains are modest and dataset-dependent. Gains over independently tuned PPR-Prob were 0.57 points on clean data and 1.09 under severe noise. Reaction OFF’s 5.31-point loss uses PtS-selected propagation parameters and does not establish the independent gain. Paired uncertainty estimates are needed to distinguish sharpening’s contribution from probability-space propagation.

We agree on Reaction OFF and on paired uncertainty. Reaction OFF keeps the (α,K)(\alpha,K) selected for PtS and only sets η=0\eta=0, so the loss of 5.31 percentage points shows how much those settings depend on sharpening, not what sharpening adds. The submission used it in this sense, see (Section 5, Table 11) of the revised paper. To avoid any ambiguity, Table 2 now contains only independently tuned differences. These are 0.57 points [0.48,0.66][0.48, 0.66] on clean inputs and 1.09 points [0.99,1.19][0.99, 1.19] at σ=2\sigma=2 for sharpening, and 1.14 points [1.04,1.25][1.04, 1.25] and 2.81 points [2.55,3.11][2.55, 3.11] for probability propagation over APPNP, with paired 95% bootstrap intervals. With all settings selected on clean validation data, sharpening still adds 0.61 points at σ=2\sigma=2 (Section 4.3). The gain is modest in the mean but not on every graph: at σ=2\sigma=2 it is 3.17 points on WikiCS and 2.72 points on Ele-Computers, both with p<0.001p < 0.001. It is dataset-dependent: positive on all nine graphs, with p<0.05p < 0.05 on five at σ=2\sigma=2 and on three on clean inputs, in paired tt-tests corrected for overlapping splits but not for multiple comparisons (Appendix J.2, Table 12).

AuthorsOpenReview ↗

Response Bxxx- Part 2/2

Global sharpening can reinforce incorrect predictions. Table 18 reports negative gains in two low-homophily bins, and PtS underperformed APPNP on severely noisy products. Sharpening preserves rankings, while Proposition 2 starts from correct predictions. The prevalence of affected nodes and whether validation can flag these failures remain unclear, limiting recommendations for general use.

Within one step, sharpening preserves each node’s class ranking, so it strengthens a wrong leading class just as it strengthens a right one. It does not freeze predictions, however: they change in the following propagation steps, and the net effect is positive. Against PPR-Prob at σ=2\sigma=2, sharpening corrects 2.8% of all test nodes and introduces errors on 1.7% (Appendix Q.1, Table 25). All nodes of a graph share one η\eta, which we state as a limitation, but sharpening is not forced on a graph. The search includes η=0\eta=0, and validation switches sharpening off in 67% of the ogbn-products units and 56% of the ogbn-arxiv units at σ=2\sigma=2 (Table 7). Proposition 2 now starts from five wrong predictions, which PtS corrects (Appendix D.2).

What fraction of each graph’s test nodes falls into Table 18’s low-homophily bins, and are negative gains concentrated in particular graphs or types of incorrect priors?

The two bins hold 14.2% of the test nodes on average, from 8.3% on Ele-Computers to 27.1% on ogbn-arxiv. Inside them PtS loses 1.12 and 0.99 percentage points against PPR-Prob, and 2.13 and 0.92 points against APPNP, which takes 0.14 and 0.17 points off the overall increase. Against PPR-Prob the losses concentrate on WikiCS, Ele-Computers, Cora-TAPE and TAPE-Arxiv23, with up to 6.17 points in the lowest WikiCS bin (Appendix Q.3, Table 27). By type of frozen prediction, over all test nodes, sharpening gains on every group of wrong predictions, 4.03 points on confidently wrong ones, and loses 0.82 points on confidently correct ones (Appendix Q.2, Table 26). Within the two low bins, the loss comes from initially correct predictions, which contribute −0.19-0.19 points to the overall gain over PPR-Prob, while initially wrong ones contribute +0.05+0.05 points.

For severely noisy products, is the degradation mainly due to propagation space, sharpening, or selection bias?

With selection at the test severity, as in Table 1, the loss comes from the propagation space: PPR-Prob is already 1.30 percentage points below APPNP, and sharpening recovers 0.09 points (Table 11). With settings selected on clean data, it comes mainly from sharpening, which accounts for 1.17 of the 1.52 points lost against APPNP. The PPR-Prob deficit remains when both methods are selected on test labels, as a diagnostic only: 0.94 points (Appendix J.3).

Can validation results flag this failure?

Partly. With validation data at the test severity, PtS is also below APPNP on validation in 41 of the 55 cases in which it is below APPNP on test, among 648 evaluations at σ=2\sigma=2, including all nine ogbn-products losses (Appendix J.4, Table 13). The other 14 losses are not flagged, and these counts do not cover clean validation under a later shift.

If output temperature is fitted only on clean validation data and then tested at σ\sigma = 2, how do PtS and APPNP compare in NLL and ECE? Could you report per-graph results, including the three graphs omitted from Table 20, to assess whether average improvements mask calibration losses?

With settings selected and the output temperature fitted on clean validation data, both applied unchanged at σ=2\sigma=2, PtS has lower mean NLL than APPNP (1.744 against 1.924) and a similar mean ECE (0.165 against 0.163). The averages do mask losses: ECE is 0.178 for PtS against 0.033 for APPNP on ogbn-arxiv, and 0.219 against 0.189 on TAPE-Arxiv23. Appendix S.1, Table 30 gives all nine graphs, including the three omitted from the original Table 20, which is now Table 29.

We hope that these answers and the added analyses resolve your questions.

Reviewer_s4TaOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies post-hoc refinement of frozen node classifiers, which takes the original categorical distribution as input and then propagate the distribution along the graph to fix the original prediction.

Its simple setting uses only the graph and the original categorical distribution. It proposes the Propagate, Then Sharpen (PtS) approach, which adds a node-wise sharpening step after the propagation. This operation strengthens existing class preferences. The design is motivated by the Dirichlet--Gini decomposition of the Potts interaction (a certain loss decomposition method stemming from CV), in which the authors make some adaptations and modifications to their final design.

They show some benign properties. The sharpening operation preserves class ordering and reduces a local KL–Gini energy (Proposition 1). On a constructed two-community graph, PtS preserves initially correct predictions at every depth, whereas APPNP and PPR-Prob with the same fixed restart parameter fail after the third iteration (Proposition 2).

Across nine homophilic graphs with a frozen MLP, "PtS improves mean test accuracy over APPNP by 1.71 percentage points on clean inputs and 3.90 under severe Gaussian feature corruption". The corresponding gains over probability propagation without sharpening (PPR-Prob) are 0.57 and 1.09 points. They also examine propagation depth, graph-aware backbones, heterophilic controls, calibration, and runtime. Code is also provided.

Critical strengths
  • The refinement method is simplistic and only concerns the graph and predicted distributions.

  • They provide a systematic evaluation, concerning sources of improvement and ablations on propagation depth, multiple backbones, transfer across corruption severities, heterophilic controls, calibration, and runtime.

Critical weaknesses
  1. The main idea is based on established propagation-and-sharpening methods.

The paper acknowledges graph-based MBO precedents, but misses several relevant works.

  • LabelRank: A Stabilized Label Propagation Algorithm for Community Detection in Networks (2013, Equation (4) and Algorithm 1) already alternates label-distribution propagation with inflation that strengthens high-probability labels.
  • GraphHop: An Enhanced Label Propagation Method for Node Classification (2021 preprint, Equation (23)) incorporates temperature sharpening into iterative graph-label learning.
  • MixMatch: A Holistic Approach to Semi-Supervised Learning (NeurIPS 2019, Equation (7)), and
  • DivideMix: Learning with Noisy Labels as Semi-supervised Learning (ICLR 2020, Equations (3)–(4)) for noisy-label refinement, already proposes sharpening after averaging predictions.

Those methods require different settings, while they have already proposed similar ideas.

  1. The experiments therefore do not establish the advantage of the proposed sharpening rule over mature alternatives proposed in the above works.

It is possible that standard temperature sharpening or hard thresholding under the same setting, with appropriate independent tuning, can achieve similar performance.

  1. The practical benefit is shown only in a restricted setting
  • The main robustness experiments only use Gaussian feature corruption, not covering the motivating cases of missing or stale features.
  • In using the GCN and GraphSAGE, the comparisons do not similarly isolate sharpening from the change in propagation space, only comparing APPNP.
  1. The significance of small incremental gains in some experiments is unknown.

Just for example, the gains over independently tuned PPR-Prob (0.57/1.09).

Questions for authors

Please just address the Critical Weaknesses above.

3 comments
AuthorsOpenReview ↗

Response s4Ta- Part 1/2

Thank you for the review, for the positive assessment of soundness, and for pointing us to the four related works. New material is marked in red in the revised PDF.

The main idea is based on established propagation-and-sharpening methods. The paper acknowledges graph-based MBO precedents, but misses several relevant works.

LabelRank: A Stabilized Label Propagation Algorithm for Community Detection in Networks (2013, Equation (4) and Algorithm 1) already alternates label-distribution propagation with inflation that strengthens high-probability labels. GraphHop: An Enhanced Label Propagation Method for Node Classification (2021 preprint, Equation (23)) incorporates temperature sharpening into iterative graph-label learning. MixMatch: A Holistic Approach to Semi-Supervised Learning (NeurIPS 2019, Equation (7)), and DivideMix: Learning with Noisy Labels as Semi-supervised Learning (ICLR 2020, Equations (3)–(4)) for noisy-label refinement, already proposes sharpening after averaging predictions. Those methods require different settings, while they have already proposed similar ideas.

We agree that alternating propagation with sharpening has precedents beyond the MBO schemes we cited, and we have added all four works to the related work (Section 1). PtS is not an MBO scheme, although it shares the alternation. It sharpens softly instead of thresholding to one-hot labels, it refines frozen predictions instead of labels, and it is based on the Dirichlet–Gini decomposition of the Potts interaction rather than the Ginzburg–Landau functional. The MBO-type hard thresholding step does worse in our controls (next answer). As you note, the four works use the idea in other settings. LabelRank is an unsupervised community-detection method, and GraphHop, MixMatch and DivideMix sharpen predictions during training. None of them refines the output of a fixed, already trained classifier from GG and QQ alone. This separates propagation and sharpening from training, which is what lets us measure each step on its own. Our contribution is an operator-splitting scheme based on the Potts decomposition in this post-hoc setting, with a mass-preserving sharpening step that is switched off exactly at η=0\eta=0 and a descent guarantee for each step (Section 2.3, Proposition 1). We add controlled evidence of what the two steps contribute: 2.81 and 1.09 percentage points of the 3.90 points over APPNP at σ=2\sigma=2 (Table 2). The sharpening operator in the equations you cite is a variant of the power map u∝pγu\propto p^{\gamma}, which is the temperature control in our next answer. (Part 1/2)

AuthorsOpenReview ↗

Response s4Ta- Part 2/2

The experiments therefore do not establish the advantage of the proposed sharpening rule over mature alternatives proposed in the above works. It is possible that standard temperature sharpening or hard thresholding under the same setting, with appropriate independent tuning, can achieve similar performance.

We ran both controls. Temperature sharpening performs similarly to our rule, and hard thresholding performs worse. With identical probability propagation, row-mass handling and off option, each tuned independently with 250 validation trials, PtS with Gini sharpening does not significantly outperform PtS with temperature sharpening: the mean differences are +0.04+0.04 percentage points (95% CI: [−0.003,0.09][-0.003, 0.09]) on clean inputs and +0.02+0.02 points (95% CI: [−0.02,0.06][-0.02, 0.06]) at σ=2\sigma=2. Both paired bootstrap intervals include zero. Hard thresholding is 0.530.53 points (95% CI: [0.45,0.60][0.45, 0.60]) below PtS on clean inputs and 0.920.92 points (95% CI: [0.79,1.13][0.79, 1.13]) below it at σ=2\sigma=2 (Section 4.2, Table 2). These results are consistent with the splitting. Our sharpening step is one majorize–minimize step on KL⁡(u∥p)+η2(1−∥u∥2)\operatorname{KL}(u \Vert p)+\tfrac{\eta}{2}(1-\lVert u\rVert^2). With the Shannon entropy H(u)H(u) in place of the Gini term, the same step on KL⁡(u∥p)+τH(u)\operatorname{KL}(u \Vert p)+\tau H(u) gives u∝p1+τu\propto p^{1+\tau}, which is temperature sharpening. Hard thresholding is the limit of both and the thresholding step of MBO schemes (new remark in Appendix B). The gain therefore comes from the soft sharpening step of the splitting, and the choice of impurity, Gini or Shannon, matters little. The paper now explicitly states that the observed gain from sharpening is not specific to the KL–Gini rule.

The practical benefit is shown only in a restricted setting

The main robustness experiments only use Gaussian feature corruption, not covering the motivating cases of missing or stale features.

The original submission only tested Gaussian corruption, so we have added whole-node feature outages, where all features of 10%, 30% or 50% of the nodes are replaced by the training means and settings are selected on clean validation data. PtS keeps a gain over APPNP at every level: 2.28 percentage points [2.08,2.49][2.08, 2.49], 2.83 points [2.35,3.31][2.35, 3.31] and 0.96 points [0.12,1.86][0.12, 1.86]. Its gain over PPR-Prob is 0.53 points [0.39,0.67][0.39, 0.67] and 0.62 points [0.32,0.89][0.32, 0.89] at 10% and 30%, and 0.01 points [−0.45,0.46][-0.45, 0.46] at 50% (Appendix M.1, Table 19). PtS itself never sees the features. It receives only QQ and GG, so the type of feature corruption matters only through the damage it does to QQ, and Gaussian noise gives a controlled ladder of that damage (Figure 3). We did not test stale features, and have removed them from the motivation.

In using the GCN and GraphSAGE, the comparisons do not similarly isolate sharpening from the change in propagation space, only comparing APPNP.

We now tune PPR-Prob independently for both backbones. At σ=2\sigma=2 sharpening adds 0.19 percentage points for GCN and 0.60 points for GraphSAGE over PPR-Prob, of the 0.88 and 1.23 points gained over APPNP, positive on all nine graphs for both, and 0.13 and 0.24 points on clean inputs (Table 3, per graph in Tables 14 and 15).

The significance of small incremental gains in some experiments is unknown. Just for example, the gains over independently tuned PPR-Prob (0.57/1.09).

For the two numbers you name, the paired 95% bootstrap intervals are [0.48,0.66][0.48, 0.66] for 0.57 percentage points and [0.99,1.19][0.99, 1.19] for 1.09 points (Table 2). Per graph the gain is positive on all nine graphs in both conditions, and paired tt-tests give p<0.05p < 0.05 on five of the nine graphs at σ=2\sigma=2 and on three on clean inputs (Appendix J.2, Table 12). The tests are corrected for overlapping splits on the six random-split graphs (Nadeau and Bengio, 2003) but not for multiple comparisons.

We hope that the new controls and the clarification of our contribution address your concerns.

Thank you for the care and effort you put into the additional experiments during the rebuttal.

My main reservation remains the size of the methodological contribution. As the revision acknowledges, earlier work already combined propagation with sharpening in related settings. In the setting studied here, the matched comparison in Table 2 shows that independently tuned temperature sharpening achieves nearly the same mean accuracy as the proposed KL--Gini rule on both clean and corrupted inputs.

I particularly appreciate the whole-node feature outage experiment, which directly addresses my concern about missing features. The new PPR-Prob comparisons for GCN and GraphSAGE also clarify how much of the gain comes from propagating probabilities and how much sharpening adds.

I can confirm my concerns are partially resolved and would maintain my original score.

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

This work focuses on post-hoc refinement for frozen models. Given only graph G and the class distribution Q output by the frozen model, with no access to features, parameters, gradients or training data, the goal is to improve node classification accuracy.

Critical strengths

Clean, solid, and honest. Negative results are fully reported (underperforms APPNP on ogbn-products, 0 accuracy on heterophilic graphs, and no gain in most cases for C&S). Limitations are well articulated; the work is rigorous with consistent experimental data. Reproduced results from the code are clean.

Critical weaknesses

Energy may be packaging, not the method itself. The abstract states that “APPNP minimizes anchored Dirichlet; we instead consider Potts energy.” However, PtS is not an algorithm for minimizing this energy. Appendix C explicitly notes that “does not imply that the PtS iteration reduces the global energy.” Moreover, Figure 6 shows that sharpening increases the Dirichlet term while only reducing the Gini term; the two terms counteract each other. More critically, regarding baselines: the stationary-point iteration for this energy is exactly LAME-Graph (Equation 9 in Appendix C), yet LAME-Graph underperforms PtS by 3.75 points at σ=2. In other words, the “canonical solution derived from the energy” is inferior to PtS. This reversely demonstrates that the actual working mechanism is the heuristic of “propagate first, then sharpen using the propagated result as anchor”, while the energy decomposition is merely a post-hoc explanation.

In the counterexample, Q itself is 100% correct (Group A: 0.9/0.1, Group B: 0.4/0.6). The search space of APPNP includes α=1, under which the output is identical to Q and achieves 100% accuracy. Hence the claim holds only for α=0.1, not for tuned APPNP. Furthermore, this example only proves that correct predictions are preserved; it does not demonstrate the correction of erroneous predictions, which differs from the paper’s main claim of “recovering accuracy from corrupted features”.

Questions for authors

I think the 2 Weaknesses should be answered.

4 comments
AuthorsOpenReview ↗

Response oZWn - Part 1/2

Thank you for the detailed review, for rerunning our code, and for describing the work as clean, solid and honest. New material is marked in red in the revised PDF. Your second weakness is correct, and we have replaced the example. On the first, we agree with the facts you cite and with part of the conclusion, and we clarify below how PtS relates to the energy.

Energy may be packaging, not the method itself. The abstract states that “APPNP minimizes anchored Dirichlet; we instead consider Potts energy.” However, PtS is not an algorithm for minimizing this energy. Appendix C explicitly notes that “does not imply that the PtS iteration reduces the global energy.”

We agree that we do not establish descent or convergence of the joint energy, and that the abstract did not state this precisely. We do not agree that the energy is only packaging: PtS is an operator-splitting scheme based on the Potts decomposition.

With the degree-normalised Dirichlet energy, the Potts interaction splits exactly into a Dirichlet part and a Gini part with the same weight at every node. For F(U)=α2∥U−Q∥F2+(1−α)E(U)F(U)=\tfrac{\alpha}{2}\lVert U-Q\rVert_F^2+(1-\alpha)E(U), propagation is a gradient step on the anchored Dirichlet part. Sharpening then applies a linearized KL-proximal step to the Gini impurity of the normalized propagated rows, before their masses are restored (Section 2.3). This follows PottsMGNet (Tai et al., 2024), which derives its layers from an operator-splitting of the Potts model. Each substep decreases its own energy: the propagation step decreases the anchored Dirichlet objective, and the sharpening step decreases the local KL–Gini energy (Proposition 1). The joint energy may increase when the substeps alternate. This is well known for operator-splitting methods, and it is what the sentence you quote from Appendix C says. It does not make the energy packaging, since its decomposition motivates both substeps. The abstract no longer says “instead we consider the Potts energy”. It now starts from the relaxed Potts interaction, whose decomposition motivates PtS.

However, the energy alone does not explain the accuracy gain. In a new control experiment, temperature sharpening matches our rule within 0.04 percentage points (Table 2). It is the same step with the Shannon entropy in place of the Gini term (Appendix B), so the gain is due to soft sharpening and not to the Gini form, and the paper now explicitly states that.

Moreover, Figure 6 shows that sharpening increases the Dirichlet term while only reducing the Gini term; the two terms counteract each other.

Yes, and this is expected in operator-splitting: each substep lowers its own part of the energy and can raise the other. Minimizing either part alone gives a degenerate solution. The pairwise Dirichlet term is minimized by identical distributions within each connected component, which is the collapse of deep propagation (Appendix D.1), and reducing Gini alone hardens each row at its current class, which changes no prediction (Appendix D.3). To assess the combined effect, we now report the total interaction energy ED+EGE_{\mathrm D}+E_{\mathrm G}, excluding the anchor. Over 100 steps, its median change is a decrease of 90.6% for PtS, compared with increases of 0.5% for APPNP and 21.8% for PPR-Prob. The total interaction energy still increases in 4.1% of the first 20 iterations of PtS (Appendix P.1, Table 23). A decrease in the total interaction energy does not by itself establish an accuracy advantage. Hard thresholding, the limit η→∞\eta\to\infty, also reduces this energy, by 89%, but is less accurate than PtS (Table 2). The sharpening strength η\eta controls the balance between the two operations and is selected by validation accuracy.

AuthorsOpenReview ↗

Response oZWn - Part 2/2

More critically, regarding baselines: the stationary-point iteration for this energy is exactly LAME-Graph (Equation 9 in Appendix C), yet LAME-Graph underperforms PtS by 3.75 points at σ\sigma=2. In other words, the “canonical solution derived from the energy” is inferior to PtS. This reversely demonstrates that the actual working mechanism is the heuristic of “propagate first, then sharpen using the propagated result as anchor”, while the energy decomposition is merely a post-hoc explanation.

We agree with both facts, and with your description of the mechanism, but not with the conclusion drawn from them.

A non-convex energy has no canonical algorithm, and we do not claim that the decomposition determines the iteration uniquely. For the graph Ginzburg–Landau functional, for example, Garcia-Cardona et al. (2014) derive both a convex-splitting scheme and an MBO scheme. LAME-Graph as evaluated is not run to an energy minimum either: its strength and number of iterations are selected by validation accuracy, as for every method (Appendix I).

The biggest difference is that LAME-Graph has no separate Dirichlet (diffusion) step, while PtS takes one in every iteration. The gradient of the Potts interaction at node ii is −(SU)i=[ui−(SU)i]+[−ui]-(SU)_i=\bigl[u_i-(SU)_i\bigr]+\bigl[-u_i\bigr], the sum of a Dirichlet part and a Gini part. LAME-Graph puts the whole Potts interaction inside one softmax around log⁡qi\log q_i, so the neighbours only add a bias to the frozen log-probabilities. PtS splits it: a gradient step on the anchored Dirichlet part, which averages neighbouring class distributions, followed by a sharpening step on the Gini part. Additionally, the fidelity terms differ. LAME-Graph keeps KL⁡(ui∥qi)\operatorname{KL}(u_i \Vert q_i) in every iteration, so the log-odds of a confidently wrong prediction are unbounded while its neighbourhood term is bounded, and the prediction is corrected only when its neighbours outweigh its log-odds. In PtS the frozen prediction enters through the restart term αqi\alpha q_i, which is bounded. We added empirical evidence consistent with this. At σ=2\sigma=2, where qiq_i is corrupted, LAME-Graph is 2.66 percentage points below plain probability propagation, which has no sharpening (Table 8). Compared with PtS, it is better on initially correct frozen predictions (90.5% against 87.4%) and worse on initially wrong ones (33.5% against 43.6%), most of all on confidently wrong ones (20.6% against 33.2%) (Table 10). The comparison therefore shows that the coupled and the split scheme behave differently under corruption, not that the split one is unrelated to the energy.

Finally, “propagate first, then sharpen using the propagated result as anchor” is not a heuristic added to the energy. In a sequential splitting each substep starts from the output of the previous one. We therefore do not agree with the statement that the “energy decomposition is merely a post-hoc explanation”. It was our starting point, and both steps are derived from it. In that sense it is an inductive bias of the method. However, our experiments do not establish why the two soft sharpening rules perform similarly or whether the Potts-based derivation explains their accuracy. Appendix C now states the two updates side by side, with both arguments.

In the counterexample, Q itself is 100% correct (Group A: 0.9/0.1, Group B: 0.4/0.6). The search space of APPNP includes α\alpha=1, under which the output is identical to Q and achieves 100% accuracy. Hence the claim holds only for α\alpha=0.1, not for tuned APPNP. Furthermore, this example only proves that correct predictions are preserved; it does not demonstrate the correction of erroneous predictions, which differs from the paper’s main claim of “recovering accuracy from corrupted features”.

Thank you for catching this. Both points are correct, and we have replaced the example (Proposition 2, Appendix D.2, Figure 5). The new example uses the same two-community graph and starts from five wrong predictions among 18 nodes. APPNP and PPR-Prob leave the same node wrong for every restart α∈[0,1]\alpha\in[0,1] and every depth, while PtS with α=0.3\alpha=0.3 and η=8\eta=8 classifies all 18 nodes correctly from depth three onward. The claim of recovering accuracy from corrupted features rests on the benchmark datasets rather than on this example: at σ=2\sigma=2, PtS corrects 23.3% of all test nodes relative to the frozen MLP and introduces errors on 5.0% (Appendix Q.1, Table 25).

We hope that these clarifications and the new example resolve both weaknesses.

Thank you for your explanation regarding the weaknesses of the paper. We believe that this work has certain innovation and quality to have a chance of being accepted by ICLR, so we are willing to give a final weakly accepted evaluation.

Thank you for your kind words and for taking the time to consider our response.

Comments on the paper

1 comment

We thank the Area Chair and the three reviewers for their careful reading and their comments, which we believe significantly strengthened the paper. All three rate the presentation as clear, and the reviews call the work “clean, solid, and honest” (oZWn), a “systematic evaluation” (s4Ta), with claims “within their demonstrated scope” (Bxxx). New material is marked in red. PtS refines a frozen node classifier from its predictions QQ and the graph GG alone. It is an operator-splitting scheme based on the Potts decomposition: a propagation step on its anchored Dirichlet part and a mass-preserving sharpening step on its Gini part, each decreasing its own energy. On nine graphs it gains 1.71 percentage points over independently tuned APPNP on clean inputs and 3.90 points under severe feature corruption.

Shared questions and changes. The main question in the reviews (s4Ta, Bxxx) is whether the gain depends on the precise KL–Gini formula. New matched controls show that it does not. Temperature sharpening matches our rule within 0.04 percentage points, while hard thresholding is 0.53 and 0.92 points worse and adds little over plain probability propagation. Temperature sharpening is the Shannon-entropy version of the same sharpening step (Appendix B), so the gain comes from the soft sharpening step of the splitting, and the choice of impurity matters little. The paper now explicitly states that the observed gain from sharpening is not specific to the KL–Gini rule. Our contribution is the splitting formulation in this post-hoc setting, with a descent guarantee for each step and sharpening that is switched off exactly at η=0\eta=0, together with controlled evidence separating probability propagation from soft sharpening. Of the 3.90-point gain over APPNP, these contribute 2.81 and 1.09 points in the nine-graph mean.

Clarifications.

  1. Energy and method (oZWn). PtS is an operator-splitting scheme based on the Potts decomposition in this post-hoc setting, in the line of PottsMGNet. The decomposition motivates both steps, with sharpening applied to the propagated distribution. Each substep decreases its own energy, while the joint energy may increase when the substeps alternate, as is typical for operator-splitting. The energy claims were already limited to local guarantees in the submission, and the abstract and Section 2.3 now state this explicitly. We agree with the reviewer that the energy alone does not explain the accuracy gain.
  2. LAME-Graph (oZWn). A non-convex energy has no canonical algorithm. LAME-Graph treats the whole Potts interaction inside one softmax and has no separate Dirichlet (diffusion) step, which PtS takes in every iteration. It also anchors to the corrupted prediction through unbounded log-odds, PtS through a bounded restart, which is consistent with its lower accuracy on initially wrong predictions (Table 10).
  3. ogbn-products (Bxxx). Under the main protocol the loss against APPNP comes from the propagation space (−1.30-1.30 percentage points) and not from sharpening (+0.09+0.09 points), as Table 10 of the submission (now Table 11) shows. With settings selected on clean data it comes mainly from sharpening, which we now report.
  4. Related work (s4Ta). The four works use sharpening in other settings, and none refines a frozen classifier. Still, the works are relevant to the broader context and we now discuss them in the introduction. The submission cited precedents for alternating smoothing and sharpening and did not claim that the alternation is new. PtS sharpens softly, and an MBO-type hard thresholding step does worse in our controls.

Corrections and additions. We replaced the Proposition 2 example that Reviewer oZWn correctly criticised, and that Reviewer Bxxx also noted, with one that starts from five wrong predictions. In the new example APPNP and PPR-Prob leave a node wrong for every restart and every depth, while PtS corrects all errors. We also added whole-node feature outages (Appendix M.1), paired uncertainty (Table 2, Appendix J.2), PPR-Prob controls for the graph backbones (Table 3), failure analyses (Appendices J.3, J.4 and Q) and calibration with a temperature fitted on clean data (Appendix S.1). The main remaining limitation is that PtS loses accuracy in low-homophily neighbourhoods (14.2% of test nodes, −0.17-0.17 percentage points on the gain over APPNP).

Apart from Proposition 2, which is now corrected, the reviews report no error in the method, proofs or experiments. Code is updated at the anonymous link.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.