Research map / Architectures and efficiency / RNNs and normalization
Beyond Pairwise Connections: Extracting High-Order Functional Brain Network Structures under Global Constraints
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GCM, a global-constraint, multi-resolution end-to-end framework that learns binary functional brain network structures from fMRI/EEG time series at sample, subject, group and project levels, using Gumbel-Sigmoid relaxation and batch-wise top-k binarization.
Problem
Functional brain network modeling relies on pairwise interactions that cannot capture high-order dependencies, and existing hypergraph approaches are computationally expensive and heuristic; FBN structures also lack explicit semantics for modeling resolution.
Approach
Treat the adjacency matrix as a single learnable object shaped by four global constraints (signal synchronization via end-to-end GNN, subject-identity contrastive loss, an adaptive expected-edge-number top-k binarization, and task labels), with explicit sample/subject/group/project modeling resolutions.
Key results
- GCM_sample reaches 0.963 intra-subject accuracy on DynHCP Activity versus 0.974 for BNT and 0.926 for BrainIB, with best inter-subject results on several datasets (e.g., 0.852 on DynHCP Activity inter).
- Subject-level GCM achieves 0.889 intra accuracy on DynHCP Age versus 0.945 for BNT at sample level, and project-level GCM gives 0.878 on DynHCP Age intra.
- Abstract claims up to 30.6% relative accuracy improvement and 96.3% reduction in computational time over baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use learned GCM structures instead of fixed-density structures (ablation) | GCM sample-level FBNs · Across reported tasks and resolutions | accuracy | Helped |
| Use adaptive BBA sparsity rather than fixed-threshold sparsity (ablation) | GCM FBNs · Across reported tasks and resolutions | accuracy | Helped |
| Use 5% fixed density rather than adaptive BBA (ablation) | GCM sample-level FBNs · DynHCP Activity Inter | accuracy: 0.837 -> 0.852 | Hurt |
| Add subject consistency contrast loss L2 (ablation) | GCM · Most conditions; group- and project-level inter-subject settings particularly | accuracy | Helped |
| Add sparsity regularization R (ablation) | GCM · Most conditions; group- and project-level inter-subject settings particularly | accuracy | Helped |
| Remove the initial Pearson-based structure (ablation) | GNN baselines · Across datasets and tasks | accuracy | Mixed |
Surprising findings
- Fully connected (dense) inputs often outperform Pearson-based FBN inputs for GNN baselines, and removing the input structure improves several GNN and GSL/BGI methods.
Limitations
- Group-level results may reflect label leakage, which the authors reinterpret rather than resolve.
- Authors note remaining room for improvement in reliable group-level FBN learning.
Experimental setup
Datasets: DynHCP Activity DynHCP Age DynHCP Gender Cog State SLIM
Benchmarks: DynHCP Activity DynHCP Age DynHCP Gender Cog State SLIM
Metrics: accuracy computation time
Baselines: LR SVM Random Forest XGBoost MLP GCN SAGE GAT GIN VIB-GSL HGP-SL DGCL
Builds on: Gumbel-Softmax (Concrete) relaxation IBGNN VIB-GSL HGP-SL BrainIB
Closest papers by meaning
- Learning Interpretable Models for Coupled Networks Under Domain Constraints (AAAI 2021 · similarity 0.90)
- Learning High-Order Relationships of Brain Regions (ICML 2024 · similarity 0.88)
- Functional Connectomes of Neural Networks (AAAI 2025 · similarity 0.87)
- $\textit{NeuroPath}$: A Neural Pathway Transformer for Joining the Dots of Human Connectomes (NeurIPS 2024 · similarity 0.86)
- A Local-Ascending-Global Learning Strategy for Brain-Computer Interface (AAAI 2024 · similarity 0.86)
- Long-range Brain Graph Transformer (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.