acceptodds
Under review as a conference paper at ICLR 2027

Cross-Scale Meta-Path Learning for Heterogeneous Brain Network Analysis

Abstract

The brain is a complex system of interconnected neurons that can be modeled as a brain network. Brain disorders often involve abnormal functional connectivity, making it important to jointly capture local interactions among brain regions and high-order coordination across functional subnetworks. However, existing brain graph learning methods mainly rely on direct pairwise connections or entangled multi-hop propagation, limiting their ability to distinguish heterogeneous connection semantics and explicitly model high-order cross-subnetwork communication. To address these issues, we propose a meta-path-guided cross-scale heterogeneous brain network learning framework. Specifically, we design a low-order heterogeneous communication module that separately models intra-subnetwork and inter-subnetwork connections through relation-specific message passing and subnetwork-aware attention. We further introduce a high-order meta-path communication module that constructs high-order communication structures within a unified functional-subnetwork meta-path space. By adaptively selecting disease-discriminative meta-paths, the module captures high-order cross-subnetwork interactions. In addition, node-level cross-scale contrastive learning is employed to align low-order and high-order representations and promote their complementarity. Experimental results on the ADNI and ABIDE datasets demonstrate that the proposed method outperforms existing approaches and identifies informative brain regions and meta-paths.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Brain Network, Meta-Path Learning, Cross-Scale Representation, Brain Disease Diagnosis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_mQ4SOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper combines relation-specific modeling of direct connections, functional-subnetwork meta-path propagation, and node-level cross-scale contrastive learning.

Critical strengths
  • Separating intra-subnetwork, inter-subnetwork, and meta-path interactions provides a reasonable inductive bias for brain network analysis.
  • Component ablations and ROI/meta-path visualizations help assess the proposed framework.
Critical weaknesses
  • Novelty over existing heterogeneous methods is unclear. BrainHGL already models functional-subnetwork meta-paths, while HeCo uses cross-view contrastive learning. Please explain the substantive advance beyond combining these established ideas.
  • Hard top-K selection raises an optimization concern. Under Equations (4)–(6), excluded path weights receive no direct gradient through the selection operation. Please explain initialization, exploration, and how previously excluded paths can become selected.
Questions for authors
  • How sensitive are results to KK, the subnetwork partition, and the contrastive weight?
  • Path-enhanced or substructure-enhanced GNNs are well-explored in graph representation learning [1-3]. Why these works are not included for discussion and comparison?

[1] Distance encoding: Design provably more powerful neural networks for graph representation learning. NeurIPS 2020 [2] Union Subgraph Neural Networks. AAAI 2024 [3] AKBR: learning adaptive kernel-based representations for graph classification. IJCAI 2025

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

The paper proposes a cross-scale heterogeneous brain network learning framework. It assigns the 90 AAL regions to seven functional subnetworks. A low-order heterogeneous communication (LHC) module models direct connections, and a high-order meta-path communication (HMC) module selects two-hop meta-paths at the subnetwork level. The two branches are trained jointly with node-level cross-scale contrastive learning. Using ten-fold cross-validation on ADNI and ABIDE, the authors report higher average performance than the compared methods on all five classification tasks.

Critical strengths
  1. The motivation is reasonable. Separating direct connections, distinguished as intra- or inter-subnetwork, from high-order meta-path communication offers a clear way to incorporate functional brain organization.

  2. The reported results are consistent, and ablations provide initial support. The method reports the highest mean ACC, F1, and AUC on all five tasks, and replacing or removing any single component reduces mean performance.

Critical weaknesses
  1. The contribution beyond existing methods is not sufficiently established. HeCo already combines a network-schema view, a meta-path view, and cross-view contrastive learning, which closely parallel LHC, HMC, and the cross-scale contrastive loss here. BrainHGL already applies meta-paths to functional subnetworks. The paper does not discuss these relationships, so it is unclear which new mechanisms account for the gains. Moreover, BrainHGL is attributed to Wen et al. (2024) in the introduction but to Xia et al. (2024) in Section 4.3, so it is unclear which baseline was actually evaluated.

  2. The central HMC design lacks targeted validation. By Equations 1–3, summing all candidate meta-path matrices yields exactly ((A^{inter})^2), so the HMC adjacency construction reweights and selects subnetwork-specific components of two-hop propagation. Removing HMC entirely does not establish that learned weighting and Top-K selection outperform simpler alternatives. It is also unclear how unselected paths receive task gradients under hard Top-K, and how the per-hop weight products in Figure 2 relate to the per-path parameters in Equations 4–6.

  3. Key methodological and evaluation details are missing. These include the AAL-to-subnetwork mapping—particularly for the hippocampus, amygdala, and thalamus highlighted in the interpretation—FC thresholding, node features, the normalization of (M), (K), the contrastive temperature, and model selection within cross-validation. ROI importance and the mapping from subnetwork-level meta-paths to the ROI connections in Figure 5 are also insufficiently defined.

Questions for authors
  1. Which specific advances over HeCo and BrainHGL account for the improvements? Which BrainHGL implementation was evaluated, and how was HeCo adapted to graph classification?

  2. How does HMC compare with uniform aggregation over all meta-paths, ((A^{inter})^2/M_p), softmax weighting without Top-K, random selection of K paths, and shuffled subnetwork labels? Please specify K, explain how hard Top-K is optimized, and clarify the relationship between Figure 2 and Equations 4–6.

  3. Could the authors provide the complete subnetwork mapping, graph construction, and model settings listed above, and explain how Figures 4–5 are generated?

  4. Are hyperparameter selection and early stopping performed entirely within each training fold? How were ABIDE subjects selected, how are sites and any repeated ADNI scans handled, and could the authors report uncertainty estimates for the paired performance differences?

Reviewer_Es2hOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a cross-scale heterogeneous brain network learning framework for neurological disorder classification. The method combines a low-order heterogeneous communication module (LHC), which separately models intra- and inter-subnetwork connectivity, with a high-order meta-path communication module (HMC), which constructs and selects cross-subnetwork paths. A node-level contrastive objective is further introduced to align the two representations.

Critical strengths
  1. Modeling local connections within brain networks separately from interactions across functional sub-networks aligns with the modular organizational characteristics of brain networks and offers a degree of interpretability.
  2. The roles of LHC, HMC, and cross-scale contrastive learning are well-defined, making the overall structure easy to understand.
Critical weaknesses
  1. High-order meta-path communication appears, in essence, to be nothing more than standard two-hop neighborhood propagation. Equation (3) calculates the product of adjacency matrices for two-hop paths—subject to constraints defined by specific sub-network types—while Equation (6) computes a weighted sum of these matrices. Consequently, although the authors emphasize that meta-paths provide unique high-order communication semantics, they fail to demonstrate that this approach can capture relationships that are unrepresentable via standard two-hop propagation.

  2. Meta-path selection lacks subject-specific adaptation, and the authors have not demonstrated that the selected paths possess stable disease semantics.

  3. Eq. (13) treats two scale representations of the same ROI within the same subject as positive samples, while labeling all other nodes, i.e, including ROIs at the same anatomical location in different subjects, as negative samples. This approach is not inherently logical in the context of brain networks, as the same brain region across different subjects may serve highly similar functional roles, yet this formulation forces them apart. The authors neither analyze this "false-negative" issue nor provide comparisons against approaches that use only intra-subject negative samples, anatomical-location-aware contrastive learning, or non-contrastive alignment methods.

  4. This paper employs ten-fold cross-validation on the ADNI and ABIDE datasets; however, it fails to demonstrate cross-site generalization capability. In the case of ABIDE, in particular, variations in multi-center data acquisition may lead the model to rely on site-specific features rather than disease-related characteristics. Furthermore, the paper does not report systematic stability tests regarding the number of selected meta-paths (KK), the network partitioning method, or the choice of ROI atlas.

Questions for authors

see the weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.