- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
Graph neural networks typically compare neighbouring messages under a shared Euclidean structure and exploit only first-order information, which cancels under averaging. This paper proposes GeoF, a recurrent framework in which each node's local propagation metric is modeled as a state that persists and evolves across recurrent steps, driven by the second moments of per-message residuals computed after frame alignment, thereby closing a feedback loop. GeoF consistently outperforms general, manifold-based, and adaptive GNN baselines on node classification, link prediction, and graph classification.
Critical strengths
-
The core idea is well targeted, and the necessity of the second order is proved. The failure mode identified is concrete and verifiable: a mean-based metric estimator assigns identical geometry to two neighbourhoods with identical means but different joint variation. The proposed remedy — per-message residuals under the current geometry, accumulated into a weighted second-moment matrix — is a direct answer, and Proposition 2 formalises the hierarchy: first-order updates cannot distinguish these configurations, the second-order target can, while 2(iv) volunteers that the target remains weaker than an injective aggregator.
-
The controls indicate that the gain comes from the mechanism rather than from extra information or capacity. Table 3 feeds the same second-order statistics to GCN/GIN/ARGNN as extra input features with parameters matched; the augmentation sometimes hurts, and GeoF still beats the best augmented variant. Table 4's
mean-only→diagonal→fullladder instantiates the moment hierarchy of Proposition 2 directly.
Critical weaknesses
-
Effect sizes are untested and capacity is uncontrolled. Several per-setting margins are at or below the baselines' own reported standard deviations, yet the paper reports only an aggregate Friedman/Holm analysis — no per-setting paired test and no confidence interval for the differences. Tables 1–2 also carry no parameter-count column, while GeoF adds per-node geometric coordinates, a controller and prototypes on top of the base model.
-
The central claim lacks an ablation that isolates it, and incrementality is not argued. Contributions (1) and (2) rest on the persistent state and the closed loop, yet no ablation localises them:
w/o GEremoves persistence, adaptation and feedback simultaneously, andw/o SFreplaces the second-order target with the current geometry rather than removing it. Meanwhile, per-node SPD metrics have precedent in ARGNN, the triangular/Cholesky parameterisation comes from Lin (2019), and SPDGNN — which uses essentially the same SPD-via-Cholesky coordinates and is itself a baseline — is never discussed in Section 2.
Questions for authors
-
(Cholesky convention). The paper is internally consistent:
Ris always lower triangular withR R^T = C + ηI, andẑis built fromR. Under theg = L^T Lreadout, however, the metric actually implemented isR^T R = R^{-1}(C+ηI)R, which shares the spectrum of the regularised residual second moment but not its principal axes. Building the target fromR^{-1}would normalise away the directions of high residual energy (whitening), whereas the current convention amplifies them; the two differ only when the residuals carry cross-dimensional coupling. Is this convention a deliberate choice, or simply inherited from the standard Cholesky factorisation? If the latter, the reading of Section 3.3 as "the target is the residual second moment" would need to be qualified, which affects my assessment of correctness. -
(Per-setting significance and parameter counts). Please report paired tests for each dataset–task setting (paired t / Wilcoxon over the 10 splits or 10 folds) together with 95% confidence intervals of the differences — especially for the settings whose margins are at or below the reported standard deviations — and add a parameter-count column to Tables 1–2. These two items directly determine the outcome of CW1 and are prerequisites for maintaining my score of 3.
-
(Effective number of prototypes). The paper reports that accuracy is nearly flat over a wide range of
K, which is compatible both with genuine insensitivity to atlas capacity and with a collapse in which only one prototype is effectively used. Please report the distribution of the effective number of prototypes per node,exp(H(α_i))withHthe Shannon entropy ofα_i, together with the fraction of nodes whoseα_iis dominated by a single prototype (e.g.max_k α_ik > 0.9). If the effective count is close to 1, the "structure-conditioned routing" component is not functioning as described and its design would need to be re-interpreted.
6 comments
Response to Reviewer ygao (2/4)
On node classification and link prediction, the confidence intervals of the paired differences exclude zero in all 12 settings, and the paired t-test gives in all 12 settings. The Wilcoxon signed-rank test agrees in 11 of them; the remaining case is link prediction on Computers (), where the confidence interval still excludes zero and the paired t-test is significant (). This holds even where the margin is close to the standard deviations reported for the baselines, because pairing on the same splits removes the variation shared by both methods, and GeoF improves on the strongest baseline consistently across splits rather than through a few favorable runs. On graph classification, GeoF improves on the strongest baseline on all six datasets, while the fold-level intervals include zero since the variance across folds on these datasets exceeds the gains. Appendix K.2 now describes these gains as consistent across datasets rather than individually significant.
Regarding capacity, GeoF uses fewer parameters than the strongest baseline in 16 of the 18 settings, often by a large margin on link prediction and graph classification. The two exceptions are node classification on PubMed and graph classification on NCI1, where the strongest baselines, AMPs and G²Former, are compact models. Capacity alone does not explain the gains in these cases either: in Section 4.6, when GCN, GIN, and ARGNN are given the same second-order statistics with parameter counts matched to GeoF within 0.5%, GeoF still outperforms them in every setting. The node-wise geometric coordinates are states computed from the shared prototype atlas and structural signatures and updated during propagation, so they are not learnable parameters and do not grow with the number of nodes. The learnable components added by GeoF are the shared controller and a small set of prototypes, and all recurrent parameters are shared across steps. The improvements therefore come from how GeoF learns propagation geometry rather than from a larger model.
W2: The central claim lacks an ablation that isolates it, and incrementality is not argued. Contributions (1) and (2) rest on the persistent state and the closed loop, yet no ablation localises them: w/o GE removes persistence, adaptation and feedback simultaneously, and w/o SF replaces the second-order target with the current geometry rather than removing it. Meanwhile, per-node SPD metrics have precedent in ARGNN, the triangular/Cholesky parameterisation comes from Lin (2019), and SPDGNN — which uses essentially the same SPD-via-Cholesky coordinates and is itself a baseline — is never discussed in Section 2.
A2: Thanks for your comments. To isolate the first two contributions, we added two variants that each change a single property while keeping the rest of the model unchanged, and report them together with the existing One-shot variant in Table 5 of Section 4.6. No persistence sets in the geometry update (Eq. 12), so each step rebuilds the geometry from the current target and correction and discards the previous state, while feedback and adaptation remain. Fixed-frame computes the residual statistics under the initial geometry while propagation still uses the current geometry, so the feedback no longer follows the geometry it updates, while persistence remains. One-shot updates the geometry only once after the first step.
| Variant | Photo | CS | Mutag | NCI1 |
|---|---|---|---|---|
| One-shot | 94.6±0.5 | 94.8±0.3 | 83.1±1.8 | 79.8±2.7 |
| No persistence | 95.0±0.5 | 95.2±0.3 | 84.1±1.4 | 82.3±2.1 |
| Fixed-frame | 95.4±0.3 | 95.6±0.3 | 84.1±1.7 | 82.6±1.8 |
| Full | 96.0±0.4 | 95.8±0.3 | 84.9±1.7 | 83.0±1.5 |
All three variants lower accuracy on all four datasets. Removing persistence causes a larger drop than fixing the feedback frame on average, and updating the geometry only once causes the largest drop on every dataset, especially on graph classification. Carrying the geometry across steps, updating it repeatedly, and computing feedback under the current geometry each contribute to performance, which localizes the persistent state and the closed loop beyond the joint effect measured by w/o GE. Regarding w/o SF, replacing the target with the current geometry reduces the update to the previous geometry plus the correction, so the second-order target is removed entirely while the correction is kept. We have clarified this in Section 4.3.
Response to Reviewer ygao (3/4)
Regarding incrementality, we have revised Section 2 to discuss SPDGNN, ARGNN, and the log-Cholesky parameterization of Lin [1]. GeoF uses this parameterization as a tool to keep the geometry positive definite during updates. SPDGNN embeds representations on the SPD manifold through Cholesky coordinates, while in GeoF the SPD geometry is a node-wise state that defines the frames used for weighting and transporting messages. ARGNN estimates node-wise metrics from features and neighborhood means at each layer, while GeoF carries the geometry as a persistent state and updates it from second-order statistics of residuals aligned under the current geometry. The No persistence variant is the closest to per-layer re-estimation, and its consistent drop shows that persistence brings improvements beyond node-wise metrics alone.
Q1: (Cholesky convention). The paper is internally consistent: is always lower triangular with , and is built from . Under the readout, however, the metric actually implemented is , which shares the spectrum of the regularised residual second moment but not its principal axes. Building the target from would normalise away the directions of high residual energy (whitening), whereas the current convention amplifies them; the two differ only when the residuals carry cross-dimensional coupling. Is this convention a deliberate choice, or simply inherited from the standard Cholesky factorisation? If the latter, the reading of Section 3.3 as "the target is the residual second moment" would need to be qualified, which affects my assessment of correctness.
A3: Thanks for your comments. The convention is defined at the level of the frame rather than the metric. In GeoF, both the neighborhood weights and the message transport are computed through the frames themselves, via and , so the object that acts on messages is and not . The target factor satisfies and maps the local unit ball onto the residual second-moment ellipsoid, so the target frame is aligned with the principal axes of the residual second moment and stretches high-energy residual directions. The induced metric shares the spectrum of but not its principal axes, as you point out. We have revised Section 3.3 and Appendix F to state that the target aligns the frame with the regularized residual second moment and that the induced metric shares its spectrum, rather than stating that the metric equals the residual second moment.
To examine whether this choice matters, we compare three conventions on Photo, CS, Mutag, and NCI1 with the rest of the model unchanged. Whitened builds the target from and normalizes away high-energy directions. Aligned builds the target so that the metric itself satisfies and therefore shares the principal axes of the residual second moment. Ours is the convention used in the paper.
| Convention | Photo | CS | Mutag | NCI1 |
|---|---|---|---|---|
| Whitened | 93.9±0.5 | 94.3±0.5 | 80.1±2.3 | 77.5±1.6 |
| Aligned | 95.7±0.4 | 95.5±0.3 | 84.3±1.7 | 82.0±1.3 |
| Ours | 96.0±0.4 | 95.8±0.3 | 84.9±1.7 | 83.0±1.5 |
Whitening consistently degrades performance on all four datasets, with the largest drops on graph classification, which shows that emphasizing rather than normalizing high-energy residual directions is essential to GeoF. Aligning the metric axes with the residual second moment performs close to our convention and slightly below it, consistent with your remark that the two differ only under cross-dimensional coupling. Together, these results indicate that the direction of the effect, amplification versus whitening, drives the benefit of second-order feedback, and that defining the target at the frame level, where it acts on messages, is a sound choice. We have added this comparison to Table 5 in Section 4.6.
Q2: (Per-setting significance and parameter counts). Please report paired tests for each dataset–task setting (paired t / Wilcoxon over the 10 splits or 10 folds) together with 95% confidence intervals of the differences — especially for the settings whose margins are at or below the reported standard deviations — and add a parameter-count column to Tables 1–2. These two items directly determine the outcome of CW1 and are prerequisites for maintaining my score of 3.
A4: Thanks for your comments. Please refer to our response to W1 (A1), which reports the per-setting paired tests, confidence intervals, and parameter counts against the strongest baseline for all 18 dataset–task settings.
Response to Reviewer ygao (1/4)
We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following:
A1: Thanks for your comments. We have replaced the aggregate Friedman and Holm analysis in Appendix K.2 with per-setting paired tests and parameter counts for all 18 settings, as shown in the table below. For each setting, GeoF is compared with the strongest baseline on the same splits or folds, and we report the mean gain, the 95% confidence interval of the paired difference, the p-values pt of the paired t-test and pW of the Wilcoxon signed-rank test, and the number of parameters. Node classification and link prediction use 10 random splits, and graph classification uses 10-fold cross-validation.