acceptodds
Under review as a conference paper at ICLR 2027

Energy-Driven Higher-Order Graph State Evolution for Brain Network Analysis

Abstract

Characterizing group-wise interactions among brain regions is important for understanding the behaviors of brain disorders. While conventional graph learning methods propagate information through pairwise edges, higher-order graph models introduce group-wise structure but commonly use it as an aggregation pattern under predefined propagation rules. Physics-inspired graph models instead derive state evolution from energy-based dynamics, but existing formulations are built from node-wise or pairwise interactions on fixed graphs. In this regard, we introduce HOPE, a higher-order graph learning framework that discovers instance-specific latent factors with sparse soft memberships and converts their group-wise interactions into factor-wise energies governing node-state evolution. HOPE alternates factor inference with structure-aware gradient-flow updates, allowing higher-order interaction structures and node states to co-evolve, while each fixed-factor state evolution follows dissipative continuous-time dynamics. Experiments on diverse brain networks show that HOPE consistently improves disease prediction across neurodegenerative and neurodevelopmental disorders. Moreover, the learned factors and energies identify disease-relevant ROI groups, revealing structured higher-order interactions associated with heterogeneous brain disorders.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Higher-order Graph Learning, Energy-based Dynamics, Brain Network Analysis

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_yXs4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

In this paper, HOPE, an energy-driven high-order graph state evolution framework, is proposed for brain network analysis. In the theoretical part, the support properties of sparse assignment of temperature control and the continuous/discrete energy dissipation properties under fixed factors are given. The experiment classifies on three brain network benchmark, and evaluates the zero-shot generalization, ablation, superparameter sensitivity, factor interpretation and efficiency comparison across acquisition phase/cohort, which is superior to baseline in many settings.

Critical strengths
  1. The modeling perspective of transforming high-order structures from aggregation operators into energy terms is essentially new. The existing high-order graph methods mostly take hyperedge/ high-order structure as the message aggregation path, while HOPE makes the learned high-order factors parameterize the total energy, and drives the state evolution through structure-aware mobility.
  2. The results of benchmark and phase/cohort zero-shot across three independent brain networks are more convincing. ADNI, PPMI and ABIDE cover neurodegeneration and neurodevelopmental disorders, and the structural/functional network modes are different.
Critical weaknesses

The conclusion of clinical classification is influenced by category imbalance and index definition, and the robustness of core promotion has not been fully proved.

Questions for authors
  1. Is F1 in Table 1 macro-F1, weighted-F1 or other definitions?
  2. Is the source domain/target domain of Zero-shot cross-phase/cohort experiment isolated by subjects? Do model selection, hyperparameters and early stopping use any labels or validation data of the target domain?
  3. Are all baseline in Table 1 retrained under the same 5-fold split, the same input features and the same parameter adjustment budget? Please give the statistical test or the result of each discount.
  4. Proposition 2 only guarantees the energy dissipation under the fixed factor. Is it possible for the total energy to rise when the alternating factor is inferred? Can you give empirical evidence of global boundedness/convergence?
2 comments
AuthorsOpenReview ↗

Response to reviewer yXs4 (Part 1)

We sincerely thank the reviewer for the constructive feedback and for recognizing the novelty of our energy-driven higher-order modeling perspective and the convincing experimental results. We address the remaining concerns, which we hope will further strengthen the reviewer's positive assessment of our work.


[W1/Q1] Class Imbalance and Evaluation Metrics

[A] We clarify that precision, recall, and F1-scores in Table 1 are macro-averaged under stratified 5-fold cross-validation, and we provide additional class-wise results using pooled out-of-fold predictions, as detailed in our response to Reviewer c1KQ [W3/Q3]. Importantly, the strong macro-F1 performance is not driven solely by majority classes, but is supported by consistent classification performance across all five diagnostic groups. Furthermore, the random-seed robustness analysis in Table 12 demonstrates that HOPE consistently outperforms strong baselines across five independent seeds. These results also support the reliability of our classification performance across different random initializations.


[Q2] Subject-level isolation and target-domain independence in zero-shot evaluation

[A] We clarify that all zero-shot experiments, including ADNI cross-phase and PPMI cross-cohort transfer, strictly maintained subject-level separation between source and target domains, ensuring that no subject appeared in both. All model training, hyperparameter tuning, model selection, and early stopping were performed exclusively using source-domain data. Neither target-domain labels nor target-domain validation data were used at any stage. Thus, the reported results demonstrate zero-shot generalization to entirely unseen subjects and cohorts without any target-specific adaptation.


[Q3] Fairness of Baseline Evaluation and Statistical Reliability

[A] We clarify the experimental protocol as follows.

  • Data Splits: All methods were evaluated using exactly the same subject-level 5-fold partitions, with a single scan per subject to prevent repeated measurements across folds. Thus, every method was compared on identical training, validation, and test subjects.
  • Input Features: All methods were provided with the same underlying brain-network information, including adjacency matrices and node features. For higher-order methods requiring incidence matrices, we constructed input hypergraphs following the strategy introduced in HGNN. Other methods, including MuHL, employed their respective architecture-specific mechanisms to construct higher-order structures from the same underlying inputs.
  • Hyperparameter Settings: Within each dataset, all methods used the same hidden width, dropout rate, learning rate, and weight decay. Specifically, the hidden widths were set to 16, 8, and 4 for ADNI, PPMI, and ABIDE, respectively, with a dropout rate of 0.5, learning rate of 10−310^{-3}, and weight decay of 5×10−45\times10^{-4}. We also employed a common hyperparameter search space and tuned model-specific parameters using validation data, as detailed in Appendix E. Furthermore, as detailed in our response to Reviewer LeHF [W2/Q2], additional hidden-width sensitivity and validation-selected capacity experiments with BQN and MuHL show that HOPE maintains accuracy advantages of 3.5%p, 3.3%p, and 2.5%p on ADNI, PPMI, and ABIDE, respectively, even when the baselines independently select their hidden dimensions using validation data, while requiring fewer parameters.
  • Statistical Reliability: We additionally evaluated ALTER, MuHL, and HOPE across five random seeds under 5-fold cross-validation, reporting mean ±\pm standard deviation and 95% confidence intervals in Table 12. HOPE maintains higher mean accuracy than MuHL across all three datasets, achieving 95.55% vs. 92.86% on ADNI, 75.95% vs. 73.37% on PPMI, and 64.18% vs. 62.86% on ABIDE. To further assess statistical significance, we conducted a two-sided exact McNemar test using paired out-of-fold predictions from HOPE and MuHL.
DatasetExact McNemar p-value
ADNI3.9 × 10⁻³
PPMI2.5 × 10⁻¹
ABIDE9.2 × 10⁻³

The results show statistically significant differences in favor of HOPE on ADNI and ABIDE (p<0.05p<0.05), while the improvement on PPMI is not statistically significant. This is understandable given the smaller sample size of PPMI (181 subjects) and the relatively modest accuracy improvement of 1.4%p. Nevertheless, HOPE maintains a higher mean accuracy than MuHL across five-seed evaluations on PPMI (75.95% vs. 73.37%), providing complementary evidence of its performance advantage. Together with the five-seed robustness results, these findings provide further support for the reliability of HOPE's performance improvements.

AuthorsOpenReview ↗

Response to reviewer yXs4 (Part 2)

[Q4] Energy Stability under Alternating Factor Re-inference

[A] We clarify that Proposition 2 guarantees continuous-time energy dissipation under fixed factors, rather than global monotonic descent across alternating factor re-inference. As detailed in our response to Reviewer LeHF [Q3] and Table 13, we separately quantified energy changes during fixed-factor state updates and factor re-inference, confirming that re-inferring the factors can increase H\mathcal{H}. To further assess the complete alternating dynamics, we extended inference to K=20K=20 without retraining. As shown in Fig. 5, H\mathcal{H} decreases sharply after initial fluctuations and remains low without sustained growth or divergent oscillations across all three datasets. Classification accuracy also remains largely stable throughout the extended evolution. As a result, these findings provide empirical support for the stability of alternating factor-state evolution over the evaluated horizon, without implying a theoretical guarantee of global convergence.

Reviewer_c1KQOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents HOPE (Higher-Order Physics-inspired Evolution), a graph learning framework designed for brain network analysis. Rather than using fixed hyperedges or standard message passing as aggregation patterns, HOPE dynamically learns instance-specific latent factors that define an energy functional, which then drives node-state evolution via gradient flow. Specifically, the model discovers soft factor assignments using a slot-based mechanism guided by both node representations and graph structural support, projecting assignment scores onto a simplex to enforce sparsity. Each factor defines an induced subgraph and an associated Dirichlet-like heterogeneity score. The total energy is defined by penalizing the squared difference between each factor's heterogeneity and a learned preferred value. Node states evolve along the negative gradient of this energy modulated by a structure-aware mobility operator (incorporating the graph Laplacian). Node states and factors co-evolve across a small number of discrete steps. The framework is evaluated across three brain network datasets (ADNI, PPMI, ABIDE) for disease classification, zero-shot cross-cohort generalization, and clinical score regression, comparing against 24 baseline models.

Critical strengths
  1. Formulating the higher-order interactions through an energy function rather than direct message passing is an interesting design choice. Squaring the factor heterogeneity naturally creates cross-interaction terms between disjoint node pairs (as shown in Appendix A and Figure 7), which provides a clear mathematical explanation for why this captures higher-order effects.
  2. The experimental section is thorough. Evaluating against 24 baseline models—spanning standard GNNs, brain-specific models, physics-inspired methods, and hypergraph networks—across three datasets gives a comprehensive picture of performance. The zero-shot evaluations across different ADNI phases and separate PD cohorts (Table 2) also test practical distribution shift scenarios.
  3. Instead of just reporting accuracy, the paper makes an effort to inspect what the model actually learns. The visualizations of factor reorganization across disease stages (Figure 3) and the perturbation experiment comparing GRAFF and HOPE (Figure 6) offer concrete intuition for how the factor-level energy behaves.
Critical weaknesses

Proposition 2 guarantees energy dissipation only under \emph{fixed} factors. However, in the proposed algorithm, factors (P(k),Q(k))(P^{(k)}, Q^{(k)}) are re-inferred at every step kk with stop-gradients on Z(k)Z^{(k)}. Once the factor memberships change, the energy landscape H(Z;P,Q)\mathcal{H}(Z; P, Q) changes as well. Figure 8 shows that energy decreases during the state update step alone, but the paper does not show whether the total energy actually decreases or stabilizes across the full alternating sequence k=0→Kk = 0 \to K. Without this, the alternating process could potentially oscillate or diverge if run for more steps.

The number of factor slots is fixed per dataset (M=20M=20 for ADNI, 10 for PPMI, 6 for ABIDE). While the sensitivity analysis in Figure 9 shows relatively stable metrics around M=20M=20 for ADNI, it is unclear what criteria should guide this choice for a new dataset. Brain networks have different numbers of functional sub-networks depending on the parcellation scheme, so hardcoding MM per benchmark limits the autonomy of the method.

In Table 4, the AD class has only 12 subjects, while CN has 226 and EMCI has 217. In a 5-fold cross-validation setup, this leaves only 2 or 3 AD subjects in the test split of each fold. Achieving 96.5% accuracy and 97.1% macro-F1 across 5 classes under such extreme imbalance raises questions about evaluation variance and potential overfitting on rare classes. The main text does not detail how this imbalance was handled in the loss function or fold splitting.

Questions for authors

In Algorithm 1, what happens to the absolute energy value H(Z(k);P(k),Q(k))\mathcal{H}(Z^{(k)}; P^{(k)}, Q^{(k)}) as kk goes from 0 to KK? Does it decrease monotonically overall, or does the re-inference of PP and QQ cause upward jumps in energy? Tracking this trajectory over more steps (e.g., K=10K=10 or 2020) would help verify whether the system converges to a stable state.

The node representations VV in Eq. 7 are defined as V=HWvV = H W_v, where H=Z+σz(A~ZWz)H = Z + \sigma_z(\tilde{A} Z W_z). When computing the energy gradient ∇ZH\nabla_Z \mathcal{H}, does the framework differentiate through the neighborhood aggregation in HH, or is HH treated as a local variable? Is the gradient calculated analytically or via automatic differentiation in PyTorch?

With only 12 AD patients out of 650 subjects, how did the authors ensure stratified splits across the 5 folds? Were any re-weighting or resampling techniques applied to the cross-entropy term in Eq. 11? Providing the confusion matrix or per-class F1-scores for ADNI would help clarify whether the high macro-F1 is driven by the majority classes.

In Figure 5, performance peaks at K=3K=3 or K=4K=4 and then levels off up to K=8K=8. If the evolution is pushed further (e.g., K=15K=15 or 2020), do the node states experience oversmoothing, or does the non-zero preferred heterogeneity ρm\rho_m prevent the representations from collapsing?

2 comments
AuthorsOpenReview ↗

Response to reviewer c1KQ (Part 1)

We really appreciate the reviewer c1KQ for the thoughtful and constructive review. We are especially thankful for the reviewer’s positive assessment, including the novelty of our higher-order energy formulation, the comprehensive empirical evaluation, and the interpretability of the learned factors. Below, we will clear out all the remaining questions.


[W1/Q1] Stability of Alternating Energy Dynamics

[A] While Proposition 2 guarantees continuous-time energy dissipation under fixed factors, we further examine the complete alternating process by extending inference to K=20K=20 without retraining, tracking absolute energy H\mathcal{H} across fixed-factor state updates and factor re-inference. To distinguish their respective effects, we additionally analyzed individual energy changes. Within the original evolution depths, fixed-factor updates exhibited no energy increases on ADNI or PPMI, while only 0.35% of updates increased energy on ABIDE. In contrast, factor re-inference increased energy in 34.92%, 12.07%, and 25.31% of updates, respectively, confirming that the alternating procedure does not guarantee monotonic energy reduction. (Detailed results are given in [Q3] for LeHF)

Importantly, Fig. 5 shows that, despite these individual fluctuations, the overall energy decreases sharply after initial fluctuations and remains low without sustained growth or divergent oscillations across all three datasets. Classification accuracy also remains largely stable throughout the extended evolution. These findings provide empirical support for the stability of alternating factor-state evolution beyond the original evolution depths, without implying a theoretical convergence guarantee for the complete alternating process.


[W2] Selection and Adaptability of Latent Factor Slots

[A] We clarify that MM controls the capacity of latent factor discovery, rather than specifying the number of anatomical or functional sub-networks. Although MM is fixed for each dataset, the factor memberships and representations are dynamically inferred for each subject and evolution step, allowing HOPE to capture instance-specific higher-order interactions without predefined ROI groupings. For a new dataset, MM can be selected through validation-based hyperparameter tuning, rather than determined by the parcellation scheme. As shown in Fig. 6, HOPE achieves its best performance around M=20M=20 on ADNI, with only modest variations at larger values, suggesting that performance does not critically depend on an exact slot count. We acknowledge that automatically determining the number of factors could further improve flexibility, and have clarified the selection criterion and this limitation in the revised manuscript.


[W3/Q3] Class Imbalance and Reliability of Classification Results

[A] We clarify that we employed stratified 5-fold cross-validation with subject-level separation, ensuring that the minority AD class was represented in every fold. No class reweighting or resampling was applied to the cross-entropy objective, and all methods were evaluated using the same data splits. To further address this concern, we conducted a class-wise evaluation using pooled out-of-fold predictions, as summarized in the tables below. Notably, HOPE correctly identifies all AD subjects, while maintaining F1-scores of at least 96.0% across the remaining classes. Together with the existing random-seed robustness analysis in Table 12, these results provide additional support for the reliability of our performance, while acknowledging the limited sample size of the AD class. Also, we clarify that all precision, recall, and F1-scores in Table 1 are macro-averaged, and we provide additional class-wise results using pooled out-of-fold predictions under stratified 5-fold cross-validation.

Table A. Class-wise classification performance on ADNI.

ClassSupportPrecision (%)Recall (%)F1 (%)
CN22694.8597.7996.30
SMC13197.6796.1896.92
EMCI21796.7495.8596.30
LMCI6498.3693.7596.00
AD12100.00100.00100.00

Table B. Confusion matrix of HOPE on ADNI. Rows represent ground-truth classes and columns represent predicted classes.

True / Pred.CNSMCEMCILMCIADSupport
CN2211400226
SMC5126000131
EMCI6220810217
LMCI10360064
AD00001212
Total2331292156112650
AuthorsOpenReview ↗

Response to reviewer c1KQ (Part 2)

[Q4] Oversmoothing and the Role of Preferred Heterogeneity

[A] As shown in Fig. 5, accuracy remains stable up to K=20K=20 across all three datasets. To further investigate representation stability and the role of preferred heterogeneity ρm\rho_m, we scale the learned ρm\rho_m by α∈{0,0.25,0.5,1}\alpha\in\{0,0.25,0.5,1\} during inference while keeping all model parameters fixed, and measure normalized node-state dispersion (DnormD_{\mathrm{norm}}) and mean pairwise cosine similarity at K=20K=20.

Experiment. Effect of preferred heterogeneity on node-state diversity at K=20K=20. Normalized dispersion Dnorm(20)D_{\mathrm{norm}}(20) measures the ratio of final to initial ROI representation variance, while Cosine(20) measures mean pairwise cosine similarity.

Datasetα\alphaDnorm(20)D_{\mathrm{norm}}(20) ↑Cosine(20) ↓
ADNI00.79310.7620
ADNI0.251.43850.6525
ADNI0.52.13750.5876
ADNI1.0 (Full)3.44160.5125
PPMI00.98740.9611
PPMI0.251.21780.9597
PPMI0.51.45040.9383
PPMI1.0 (Full)1.82250.9252
ABIDE00.93080.9512
ABIDE0.252.22380.8832
ABIDE0.53.57870.8586
ABIDE1.0 (Full)6.90800.8324

As shown above, increasing α\alpha consistently increases normalized node-state dispersion and reduces cosine similarity across all three datasets, indicating that learned preferred heterogeneity promotes more differentiated ROI representations. In particular, full HOPE maintains nonzero representation differences at K=20K=20, with final dispersion exceeding the initial level across all datasets. Together with the stable classification accuracy in Fig. 5, these results provide empirical support for the role of preferred heterogeneity in maintaining representation diversity during extended evolution.

Reviewer_LeHfOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

HOPE infers sparse subject-specific latent factors from node features and graph structure, uses these factors to define nonlinear energies, and updates node states through a structure-aware gradient flow. Factor inference and state evolution alternate, with a fixed-factor dissipation result and an analysis of temperature-controlled assignment sparsity. Experiments cover ADNI, PPMI, and ABIDE, including cross-phase or cross-cohort transfer, component ablations, and additional robustness analyses.

Critical strengths

The energy construction provides a concrete modeling contribution. Appendix A shows how applying a nonlinear potential after aggregating factor-wise discrepancies creates cross-interaction terms, including between disjoint edges; the higher-order claim is therefore supported by an explicit derivation rather than by terminology alone.

The empirical study is broad and relevant to the claimed contribution: it includes 24 baselines, physics-inspired alternatives, fixed-factor and conventional-energy ablations, transfer experiments, and repeated-seed results for strong competitors. The paper also correctly limits its dissipation guarantee to fixed-factor updates instead of claiming a global Lyapunov result for the entire alternating procedure.

Critical weaknesses

The energy ablation changes two things at once. Replacing the proposed potential with conventional Dirichlet energy removes the cross-interaction terms but also removes the learned preference for nonzero heterogeneity. The observed drop therefore supports the combined design, but does not establish that nonseparable higher-order coupling is the decisive ingredient.

The baseline-capacity settings weaken the comparison on the functional-network datasets. Appendix E gives baseline hidden widths of 8 on PPMI and 4 on ABIDE, while HOPE uses a latent width of 16 throughout. The extensive baseline list is useful, but the strongest competitors should be allowed comparable validation-selected capacity before the performance margin is treated as settled.

Questions for authors

Can the cross-interaction effect be isolated with a separable nonlinear edge-energy control that retains the same factor assignments and a comparable preferred-heterogeneity parameterization? This would resolve the main ambiguity in the conventional-Dirichlet ablation.

Are the reported baseline widths fixed as described in Appendix E, or selected individually from the search grid? Please report the strongest baselines with independently validation-selected widths and the corresponding parameter counts.

Figure 8 supports fixed-factor dissipation at the aggregate level. With a fixed Euler step of 0.25 and learned mobility coefficients, how often does an individual update increase energy, and how large are the changes introduced by re-inferring the factor?

What label mapping is used for transfer from the three-class PPMI task to TaoWu and Neurocon? Please specify whether prodromal subjects are excluded, merged with another class, or handled by collapsing predictions, and whether the same rule is used for all methods.

Please reconcile the CDR-SOB metrics in Table 9: HOPE has higher RMSE than MuHL but also higher R-squared. Are these metrics averaged separately across folds or computed from pooled predictions? Consistent out-of-fold metrics would resolve this ambiguity.

8 comments
AuthorsOpenReview ↗

Response to reviewer LeHf (Part 1)

We sincerely thank Reviewer LeHF for the constructive feedback and for recognizing the theoretical novelty of our higher-order coupling mechanism and the breadth of our empirical evaluation. Below, we address the remaining concerns and hope that the clarifications will positively influence the reviewer’s assessment toward acceptance.


[W1/Q1] Isolating the contribution of higher-order energy coupling

[A] We appreciate the reviewer's insightful suggestion to further isolate the contribution of higher-order coupling. To address this concern, we additionally introduce a separable nonlinear energy variant that retains the same factor assignment mechanism and preferred-heterogeneity parameterization while removing the cross-interaction terms between distinct edges.

MethodCross-InteractionPreferred HeterogeneityADNI AccADNI PreADNI RecABIDE AccABIDE PreABIDE Rec
w/o H\mathcal{H}✗✗93.9±1.995.0±1.294.1±1.361.7±1.961.8±2.061.3±2.1
w/o cross-interaction✗✓94.4±2.295.6±2.195.0±1.663.4±2.463.5±2.763.0±2.8
HOPE (Full)✓✓96.5±2.297.5±1.696.7±1.565.7±2.465.7±2.365.0±2.9

Here, the separable variant improves accuracy over conventional Dirichlet energy (w/o H\mathcal{H}) by 0.5%p on ADNI and 1.7%p on ABIDE, which confirms the benefit of learned preferred heterogeneity. More importantly, the full HOPE further improves accuracy by 2.1%p and 2.3%p, respectively, demonstrating the additional contribution of nonseparable higher-order coupling. These results clarify that the performance gains arise not merely from learned heterogeneity preferences, but also from the proposed cross-interactions between distinct edges.

AuthorsOpenReview ↗

Response to reviewer LeHf (Part 2)

[W2/Q2] Fairness of baseline capacity settings

[A] We clarify that HOPE and all baselines originally used the same dataset-specific hidden dimensions, namely 16, 8, and 4 for ADNI, PPMI, and ABIDE, respectively. The description in Appendix E incorrectly suggested that HOPE used a hidden dimension of 16 throughout, and we have corrected this inconsistency. To further address this concern, we conducted two additional experiments with BQN and MuHL, representative brain-specific and higher-order baselines, respectively. First, we evaluated all three methods across the same hidden-width grid, d∈{2,4,8,16,32,64}d \in \{2,4,8,16,32,64\}, as shown in Experiment A. The results show that HOPE achieves the highest accuracy on all three datasets across the explored widths, indicating that its performance advantage is not simply attributable to a larger hidden dimension.

Experiment A. Hidden-width sensitivity of HOPE and the strongest baselines across three datasets. Bold values indicate the original dataset-specific hidden dimensions used in Table 1.

DatasetMethodd=2d=4d=8d=16d=32d=64
ADNIBQN78.479.480.381.580.981.0
ADNIMuHL92.192.092.693.292.892.4
ADNIHOPE93.095.395.596.595.494.8
PPMIBQN60.369.071.968.368.965.3
PPMIMuHL65.571.276.868.570.167.9
PPMIHOPE69.076.478.272.576.473.8
ABIDEBQN61.363.362.162.961.360.9
ABIDEMuHL62.163.763.263.863.761.5
ABIDEHOPE64.465.765.163.862.262.0

Second, we independently selected the hidden dimension for each baseline using validation accuracy within each fold, without using held-out test data for model selection. As shown in Experiment B, the validation-selected BQN achieves accuracies of 82.0%, 70.0%, and 62.7%, while MuHL achieves 93.0%, 74.9%, and 63.2% on ADNI, PPMI, and ABIDE, respectively. HOPE retains accuracies of 96.5%, 78.2%, and 65.7% under the original shared-width settings, outperforming the strongest validation-selected baseline by 3.5%p, 3.3%p, and 2.5%p, respectively. Notably, HOPE also uses fewer parameters than both validation-selected baselines across all three datasets, further supporting that its improvements are not merely due to greater model capacity.

Experiment B. Classification accuracy with independently validation-selected hidden dimensions. The selected widths are reported for each fold, while HOPE retains its original dataset-specific fixed width. Parameter counts are averaged across the five fold-specific configurations.

DatasetMethodHidden Widths (5 Folds)Acc (%)Params (M)
ADNIBQN32, 16, 4, 16, 482.05.279
ADNIMuHL64, 16, 2, 2, 1693.031.577
ADNIHOPE16 (fixed)96.53.313
PPMIBQN32, 2, 8, 4, 870.02.639
PPMIMuHL8, 32, 4, 8, 3274.96.433
PPMIHOPE8 (fixed)78.20.437
ABIDEBQN4, 8, 64, 2, 462.70.942
ABIDEMuHL4, 2, 2, 32, 463.23.550
ABIDEHOPE4 (fixed)65.70.109

Both experiments demonstrate that HOPE maintains its performance advantage under both shared hidden-width settings and independently validation-selected baseline capacities, while requiring fewer parameters. These findings further substantiate the fairness of our comparisons and show that the observed improvements cannot be explained solely by differences in hidden dimensions or parameter counts.

AuthorsOpenReview ↗

Response to reviewer LeHf (Part 3)

[Q3] Energy Dissipation under Individual State Updates and Factor Re-inference

[A] To assess individual energy changes, we measured H\mathcal{H} before and after each fixed-factor state update using the original Euler step size (Δt=0.25\Delta t=0.25) and learned mobility coefficients across all held-out subjects. Within the original evolution depths, no energy increases occurred on ADNI or PPMI, and only 0.35% of updates increased energy on ABIDE. Extending inference to K=20K=20 revealed occasional increases beyond the trained horizon, as summarized below.

Experiment A. Energy variations during fixed-factor state update ((P,Q)(P, Q): fixed, ZZ: updated). Mean and maximum positive jumps represent relative energy increases among energy-increasing updates.

DatasetTrained Increase (%)Extended Increase (%)Mean Positive Jump (%)Max Positive Jump (%)
ADNI0.00 (0/1950)10.4523.8773.07
PPMI0.00 (0/543)0.00——
ABIDE0.35 (14/3956)10.3129.90129.40

We further quantified energy changes caused by factor re-inference. Re-inferring (P,Q)(P,Q) increased energy in 34.92%, 12.07%, and 25.31% of updates on ADNI, PPMI, and ABIDE, respectively, with mean signed energy changes of +0.2792, −0.0001, and +0.8507. These changes arise from re-inferring the factors that define H\mathcal{H} and are not covered by Proposition 2, which establishes continuous-time dissipation under fixed factors.

Experiment B. Energy variations during factor re-inference ((P,Q)(P, Q): updated, ZZ: fixed).

DatasetFactor Re-inference Increase (%)Mean Signed Energy Change
ADNI34.92+0.2792
PPMI12.07−0.0001
ABIDE25.31+0.8507

Furthermore, Fig. 5 tracks absolute energy and classification accuracy over 20 evolution steps across all three datasets. Despite occasional increases caused by factor re-inference, the overall energy decreases rapidly after initial fluctuations and remains low without sustained growth or divergent oscillations. Also, classification accuracy remains largely stable beyond the original evolution depths. Overall, these findings support the empirical stability of alternating factor-state evolution while clarifying that the theoretical dissipation guarantee applies only to continuous-time dynamics under fixed factors.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Answer

We sincerely thank all reviewers for their constructive comments and positive assessments of our work. We especially appreciate their recognition of our core novelty in formulating learned higher-order interactions as energies that directly govern graph state evolution, rather than merely serving as aggregation operators.

In the following, we carefully address all raised concerns and questions, and the corresponding revisions are highlighted in red\color{red}{\text{red}} in the updated manuscript. We hope that our clarifications and additional analyses will resolve the remaining concerns and further strengthen the reviewers' positive assessments toward acceptance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.