acceptodds
Under review as a conference paper at ICLR 2027

Multi-Hop Signed Graph Convolutional Network Under Long-term Dependencies Supervision for Neurodevelopmental Disorder Diagnosis

Abstract

Brain network analysis leveraging resting-state functional magnetic resonance imaging offers promising prospects for the diagnosis of neurodevelopmental disorders (NDDs). Graph neural networks have demonstrated their powerful ability to extract informative latent features to facilitate diagnostic analysis. Nevertheless, existing signed multi-hop graph convolutional networks (GCNs) are plagued by insufficient prior supervision and the over-smoothing problem, which degrades diagnostic performance. In this study, we propose a long-term guided multi-hop signed graph convolutional network (LT-MSGCN) to identify autism spectrum disorder (ASD) and attention deficit hyperactivity disorder (ADHD). We construct a dual-branch architecture consisting of an auto-encoding (AE) branch and a multi-hop GCN branch. Specifically, the transformer-augmented AE branch captures long-term dependencies among brain regions, and the resulting relationship matrices serve as regularizers for multi-hop adjacency matrices generated via approximate personalized propagation of neural predictions (APPNP) within the multi-hop GCN branch. After feature extraction by stacked signed graph convolution layers, hop-aware and region-adaptive attention fusion modules further boost discriminative graph representation learning. Under 10-fold cross-validation on the ABIDE-I and ADHD-200 datasets, LT-MSGCN achieves classification accuracies of 71.53% and 70.20%, respectively. The identified discriminative brain regions show good agreement with established neuroscientific findings. Experimental results demonstrate that LT-MSGCN yields competitive classification performance and favorable biological interpretability for the auxiliary diagnosis of NDDs.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
APPNP, functional connectivity, multi-hop, neurodevelopmental disorder, signed graph convolution

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_aEJMOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes LT-MSGCN, a multi-hop signed graph convolutional network incorporating long-term dependency supervision for neurodevelopmental disorder classification, with a focus on Autism Spectrum Disorder (ASD) and Attention Deficit Hyperactivity Disorder (ADHD).

To address the limitations of conventional GCNs in modeling negative functional connectivity, long-range interactions between brain regions, and over-smoothing, the authors introduce a dual-branch architecture. A Transformer-based autoencoder (AE) learns long-term relation matrices between brain regions, while a multi-hop GCN branch constructs hop-specific adjacency matrices using APPNP and extracts graph representations through Signed GCN. The two branches are jointly optimized with a consistency loss between the learned relation matrices and multi-hop adjacency matrices. The model also incorporates hop-aware and region-aware feature fusion modules.

The authors evaluate the model using 10-fold cross-validation on ABIDE-I and ADHD-200, reporting classification accuracies of 71.53% and 70.20%, respectively. They further investigate biological interpretability through between-group statistical analyses of learned brain-region relationships.

Critical strengths
  1. The research problem is well motivated.

Modeling both positive and negative functional connectivity, together with long-range interactions between brain regions, is relevant to brain disorder classification. The authors identify limitations of conventional GCNs in local information aggregation and the lack of explicit guidance for multi-hop propagation, providing a reasonable motivation for the proposed approach.

  1. The proposed framework is relatively comprehensive.

The paper integrates multi-hop graph propagation, relation modeling, and attention-based feature fusion into a unified framework. It also includes comparative experiments, component ablations, and biological analyses on two disorder classification tasks, demonstrating potential practical relevance.

Critical weaknesses

W1. Limited Methodological Novelty and Insufficient Validation of the Core Contribution

The main components of the proposed method, including APPNP-based multi-hop propagation, Signed GCN, Transformer-based relation modeling, and hop-wise attention, are largely based on established techniques. The primary proposed contribution appears to be the use of Transformer AE to learn relation matrices and guide multi-hop propagation through a consistency loss. However, the novelty and necessity of this mechanism are not sufficiently established.

Table 2 shows that removing the AE branch reduces accuracy by only 0.88 and 0.19 percentage points on ABIDE-I and ADHD-200, respectively, whereas removing the multi-hop component leads to decreases of 5.92 and 5.75 percentage points. This suggests that much of the performance gain may come from the existing multi-hop propagation mechanism. Moreover, the paper does not independently ablate the guidance loss or disentangle the contributions of AE reconstruction and relation alignment, leaving the role of long-term dependency supervision insufficiently validated.

W2. Unclear Model Selection Protocol and Limited Performance Improvements

The paper adopts 10-fold cross-validation but does not specify whether a separate validation set is used or how the final model checkpoint is selected. Section 3.3 identifies the optimal number of hops based on classification performance without clarifying which data are used for hyperparameter selection. Consequently, potential test-set selection bias cannot be ruled out.

Despite introducing multiple architectural components, the performance improvements are limited and inconsistent. According to Table 1, the proposed model does not achieve state-of-the-art accuracy or AUC on ABIDE-I, while its accuracy improvement over the strongest competing baseline on ADHD-200 is only 0.60 percentage points. Given the model's complexity, these results do not establish a clear and consistent performance advantage. Furthermore, the absence of standard deviations, confidence intervals, or significance tests, together with insufficient clarification of baseline evaluation protocols, weakens the reliability of the reported comparisons.

W3. Mathematical Inconsistencies and Insufficient Justification of the Propagation Mechanism

Section 2.4 claims that if all elements of the original adjacency matrix lie within (-1, 1), the resulting multi-hop matrices A_i must also lie within this range. However, this claim is mathematically incorrect. For example, consider a three-node adjacency matrix with zero diagonal entries and all off-diagonal entries equal to 0.9. With α = 0.1, Equation (5) yields (A₂)₁₁ = 1.4122, which exceeds 1. This raises concerns about a potential scale mismatch between A_i and the tanh-bounded relation matrices R_i. Moreover, the paper does not adequately clarify whether the normalization assumptions and spectral properties underlying standard APPNP remain applicable to the proposed signed multi-hop propagation.

In addition, the Frobenius-norm loss in Equation (14) imposes a consistency constraint between R_i and A_i, but does not by itself demonstrate that the learned Transformer relations effectively guide multi-hop propagation. The paper also lacks direct theoretical or empirical evidence that the proposed mechanism alleviates over-smoothing. Therefore, the mathematical properties, direction of supervision, and actual contribution of the core propagation mechanism remain insufficiently justified.

W4. Inadequate Biological Validation and Insufficient Support for Interpretability Claims

Section 4 and Appendix C identify disease-associated connections through between-group t-tests on the learned relation matrices R_i. However, the paper does not clearly state whether multiple-comparison correction was applied or how potential confounding factors, such as age, sex, acquisition site, and head motion, were controlled. Moreover, since R_i is obtained from a supervised classification model, the lack of clarity regarding which samples were used for statistical testing raises concerns about potential circular statistical inference.

The authors further interpret the decreasing number of significant connections in higher-order relation matrices as evidence of more stable and disease-specific patterns. However, this interpretation is not supported by cross-fold stability analysis, independent validation, or an assessment of the identified connections' contributions to model predictions. Therefore, the current evidence is insufficient to establish these connections as reliable disease-specific biomarkers.

W5. Inconsistent Reporting and Insufficient Reproducibility Details

The paper contains several reporting inconsistencies and missing implementation details. The Contributions section reports classification accuracies of 70.2% and 71.8%, which differ from the 71.53% and 70.20% reported in the Abstract and Table 1. In addition, important methodological details remain unclear, including the handling of self-loops, constraints on APPNP parameters, aggregation of evaluation metrics across folds, and the mapping from AAL90 ROIs to Yeo7 functional networks. These omissions hinder reproducibility and independent verification of the reported results.

The presentation of biological findings also requires improvement. Figures 3 and 4 do not clearly distinguish statistical significance, learned relation weights, brain-region importance, and functional network assignments. The criteria used to identify and rank important brain regions are insufficiently explained, making the results difficult to interpret. These presentation issues further limit the clarity and reproducibility of the biological analyses.

Questions for authors

Q1. Model Selection and Validation

Could the authors clarify the model selection procedure within the 10-fold cross-validation framework? Was a separate validation set used within each fold? How were the final checkpoints, the optimal number of hops in Section 3.3, and other hyperparameters selected? Were the outer test folds used at any stage of model selection?

Q2. Contribution of Long-Term Guidance

Could the authors provide targeted ablations that independently remove the guidance loss, retain the AE reconstruction objective, or replace the Transformer relation matrices with simpler alternatives? What evidence demonstrates that R_i provides more effective guidance than the existing APPNP propagation matrices?

Q3. Fairness and Reliability of Experimental Comparisons

Were all baselines in Table 1 evaluated using consistent data splits, preprocessing settings, and training protocols? For baselines whose settings or results differ from their original publications, could the authors provide reimplementation details? Could the authors also report variability across runs or confidence intervals to establish the reliability of the relatively modest performance gains?

Q4. Mathematical and Implementation Details

Was the original adjacency matrix normalized before computing its higher-order powers? How is the numerical range of the multi-hop matrices controlled? Was the adjacency matrix symmetrized after per-node top-20% sparsification? How do gradients from Equation (14) propagate through the AE and APPNP branches?

Q5. Reliability of Biological Validation

Which samples and model checkpoints were used to obtain the R_i matrices for the biological analyses? Were multiple-comparison correction and confound control performed? Could the authors provide cross-fold stability analyses or independent validation to support the reliability and disease specificity of the identified connections?

1 comment

I have carefully revisited my original review and confirmed that the concerns raised remain valid. As the authors have not provided a response addressing these concerns, I will maintain my original rating.

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

This paper proposes a LT-MSGCN for neurodevelopmental disorder diagnosis. Specifically, an autoencoder is used to learn relationship matrices among brain regions, which are then used to regularize multi-hop adjacency matrices constructed through APPNP-style propagation. On this basis, signed graph convolution and adaptive feature fusion modules are further introduced to extract discriminative brain representations. Finally, experimental results demonstrate the competitive classification performance of LT-MSGCN, accompanied by brain-region-level analyses to provide biological interpretability.

Critical strengths
  1. The paper addresses a meaningful problem in neurodevelopmental disorder diagnosis by exploring the modeling of positive and negative functional connectivity and multi-hop interactions in brain networks.

  2. LT-MSGCN integrates long-range relationship modeling, multi-hop graph propagation, and adaptive feature fusion. Experiments on two publicly neuroimaging datasets, together with component ablation studies and brain-region-level analyses, provide preliminary evidence supporting the effectiveness.

Critical weaknesses
  1. What is the fundamental difference between the proposed SGCN and a conventional GCN?

  2. Could the authors provide a mathematical proof that the elements of the multi-hop adjacency matrices always remain within the interval (-1,1)?

  3. Why can the attention matrices generated by Transformer be regarded as reliable supervision signals for long-term dependencies?

  4. What fundamentally distinguishes this mechanism from conventional cross-branch consistency regularization?

  5. Ccould the authors validate the independent contribution of the proposed Long-term Dependency Supervision mechanism?

Questions for authors

See Weaknesses

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

This paper proposes LT-MSGCN for binary classification of autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD) using resting-state fMRI functional connectivity graphs. The method combines a transformer-based auto-encoding branch that learns long-term relationship matrices with a multi-hop graph convolutional branch that captures higher-order connectivity patterns. A guiding loss aligns the learned attention matrices with corresponding multi-hop adjacency matrices, facilitating the integration of long-range relationships into graph aggregation along with cross entropy and reconstruction loss. Evaluated using 10-fold cross-validation on ABIDE-I and ADHD-200, the method achieves accuracies of 71.53% and 70.20%, respectively. The paper also presents comparisons with existing GNN methods, ablation studies, hop-number analysis, and attention-based investigations of potential neuroimaging biomarkers.

Critical strengths

S1. Coherent architectural motivation. The paper guides multi-hop graph aggregation using long-range relationships learned by a transformer encoder, providing a principled motivation for integrating the two components. Additionally, the signed graph convolution preserves positive and negative functional connectivity values through absolute-degree normalization, retaining potentially informative anti-correlations.

S2. Comprehensive component-level evaluation. The authors conduct ablation studies covering the major architectural components (multi-hop aggregation, the auto-encoding branch, signed convolution, and both fusion modules). Removing multi-hop aggregation reduces accuracy from 71.53% to 65.61% on ABIDE-I and from 70.20% to 64.45% on ADHD-200. The hop-number analysis further supports the effectiveness of four-hop aggregation on both datasets.

S3. Evaluation on two public datasets. The method is evaluated on ABIDE-I and ADHD-200 for ASD and ADHD classification, respectively. The paper reports dataset demographics, architectural configurations, and training settings, providing useful experimental details and evidence across two neuro-imaging classification tasks.

S4. Competitive performance on ADHD-200. LT-MSGCN achieves the highest reported accuracy (70.20%) and sensitivity (75.55%) among the compared methods on ADHD-200. It also achieves an AUC of 70.34%, outperforming the reported BrainOSM result of 66.52%. These results indicate competitive performance on this dataset.

Critical weaknesses

W1. Validity of the range assumption for multi-hop adjacency matrices. The APPNP-style expansion in Eq. 4 appears reasonable; however, the claim that every A_i remains within (-1,1) (p. 5, lines 242–243) may not be generally guaranteed, as entries of matrix powers aggregate products over multiple walks and can exceed this range. Although the normalization in Eq. 6 may mitigate this issue during graph convolution, the guiding loss in Eq. 14 directly compares unnormalized A_i with tanh-bounded R_i. Consequently, the target may be unattainable, and the loss scale may vary with the hop number and \alpha. Please clarify the validity of this assumption and discuss its implications for the guiding loss.

W2. Insufficient statistical support for the reported improvements. Tables 1–2 report mean results without significance tests, and the error bars in Fig. 2 require more info. This is particularly important for small performance differences: removing the auto-encoding branch reduces accuracy by only 0.88 percentage points on ABIDE-I and 0.19 percentage points on ADHD-200, while the improvement over RostGNN on ADHD-200 is only 0.60 percentage points. Furthermore, Eq. 4 suggests that \alpha is the only learnable parameter in the multi-hop adjacency expansion. The paper does not sufficiently establish how the guiding loss contributes beyond optimizing this parameter and the shared input representations. Please clarify the statistical reliability of these improvements and provide a more explicit analysis of the guiding loss’s contribution.

W3. Overstated comparative claims. The contributions section claims that the experiments verify the superiority of LT-MSGCN (p. 3, lines 122–123), but the reported results do not consistently support this conclusion. On ABIDE-I, BrainOOD achieves higher accuracy (72.06% versus 71.53%), and four baselines report higher AUC values than LT-MSGCN. The proposed method does not achieve the best result on any reported ABIDE-I metric, while its specificity on ADHD-200 is the lowest among the compared methods (62.65%). Although Section 3.1 acknowledges BrainOOD’s advantage, the broader claims of superiority should be qualified. Additionally, given the use of different brain atlases, it is unclear whether all baselines were evaluated on identical subjects and cross-validation folds. Please clarify the evaluation protocol across baselines and revise the comparative claims to reflect the reported results more accurately.

W4. Potential demographic and site-related confounding. Table 3 indicates a substantial gender imbalance between the ADHD (198 female, 45 male) and typically developing (163 female, 135 male) groups in ADHD-200. The use of random subject-level splits on multisite data, without clearly described stratification, covariate adjustment, site harmonization, or site-held-out evaluation, raises concerns about potential demographic and site-related confounding. These factors may limit the generalizability of the reported results and the claim of generalization capability (p. 3, line 125; p. 7, lines 359–360). Please clarify how these factors are addressed.

W5. Limited novelty and broader methodological significance. Most components of LT-MSGCN, including APPNP-style propagation, transformer-based autoencoding, signed graph convolution, and attention-based fusion, are established techniques. The guiding loss appears to be the primary distinctive contribution. However, Appendix A discusses related approaches, including learnable APPNP weights in MCMGN and multi-hop signed aggregation in MOSGCN, without clearly establishing the technical distinction from the proposed method. Furthermore, evaluation on only two functional-connectivity datasets leaves the broader methodological significance and applicability of the approach insufficiently demonstrated. Please articulate the precise novelty of the guiding loss relative to existing methods.

Minor / Presentation Issues

W6. Inconsistent performance figures. The contributions section reports accuracies of 70.2% and 71.8% for ABIDE-I and ADHD-200, respectively (p. 3, lines 123–124), conflicting with the abstract and Table 1, which report 71.53% and 70.20%. These discrepancies should be corrected.

W7. Notational and referencing issues. The symbol \alpha is used for both the APPNP propagation weight in Eq. 4 and the loss weights in Eq. 15, potentially causing confusion. The parameter d_k is not defined, and MixHop appears to be cited twice.

Questions for authors

Q1. Statistical significance and variability: Could the authors report mean ± standard deviation across folds, or results from repeated cross-validation with different random seeds, for Tables 1 and 2? Could you also provide paired statistical tests or confidence intervals comparing LT-MSGCN with the strongest baseline and the full model with Variant 2? Additionally, what do the error bars in Fig. 2 represent?

Q2. Guiding loss and multi-hop adjacency construction: How is the parameter \alpha parameterized? Is it shared across hops, and is it constrained to (0,1)? Could the authors report its learned values and clarify whether A_i is normalized before being used in the guiding loss \mathcal{L}_{hop}? Furthermore, could an ablation study removing only the guiding loss, while retaining the auto-encoding branch and the rest of the architecture, be provided to understand its contribution?

Q3. Generalization and demographic confounding: Could the authors report per-site performance or leave-one-site-out evaluation across the four ADHD-200 sites to assess generalization to unseen sites? Additionally, could you report the performance of a classifier using only demographic variables, particularly gender, and site information? Finally, could the gender counts reported in Table 3 be confirmed?

Q4. Fairness of baseline comparisons and methodological novelty: Were all baselines in Table 1 evaluated using the same subjects and cross-validation folds, given the differences in brain atlases used across methods? Could the authors clarify the experimental protocol to ensure a fair comparison? Furthermore, how does the proposed guiding loss differ conceptually and technically from the regularized learnable APPNP weights used in MCMGN?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.