| name | hyperbolic-learning-brain-graphs-hlbg |
| description | Hyperbolic Learning on Brain Graphs (HLBG) methodology for brain disorder diagnosis using Lorentzian hyperbolic space and Graph-aware Mamba (GaMamba). Models hierarchical relationships among ROIs, functional communities, and whole-brain networks via geometric entailment constraints. Activation: hyperbolic brain graphs, brain network diagnosis, GaMamba, hyperbolic space brain, Lorentz model brain, hierarchical brain representation.
|
Hyperbolic Learning on Brain Graphs (HLBG)
From: "Navigating Hierarchy: Hyperbolic Learning on Brain Graphs for Disorder Diagnosis" (arXiv:2607.07077, July 2026) — Yapeng Li, Bo Jiang, Ziyan Zhang, Dongdong Chen, Zhengzheng Tu.
Core Concept
Functional brain networks exhibit hierarchical organization across ROI, community, and whole-brain levels. HLBG exploits the inherent hierarchical geometry of hyperbolic space (Lorentz model) to model these relationships, learning hierarchy-aware and highly discriminative representations for brain network analysis.
Problem
Existing brain graph modeling methods struggle to model ROI–community interactions, failing to fully exploit the hierarchy across ROI, community, and whole-brain network levels.
Solution
Project representations from ROIs, communities, and whole-brain network into Lorentzian hyperbolic space, then impose multi-level hierarchy via geometric entailment constraints.
Architecture
1. Hierarchical Brain Graph Construction
- Build graphs at three levels:
- ROI level: individual brain regions as nodes
- Community level: functional communities as nodes
- Whole-brain level: global network
- Each level captures different scales of brain organization
2. Graph-aware Mamba (GaMamba)
- Novel contribution: incorporates topology-derived structural prompts into Mamba architecture
- Captures long-range dependencies while preserving graph topological information
- Mamba's state-space model adapted for graph-structured data
3. Hyperbolic Projection
- Project representations from all three levels into Lorentzian hyperbolic space
- Hyperbolic space naturally encodes hierarchical structure with minimal distortion
- Uses the Lorentz model (not Poincaré ball) for numerical stability
4. Geometric Entailment Constraints
- Two geometric entailment constraints impose the multi-level hierarchy:
- ROI → Community entailment: ROI representations are geometrically "contained" within their community representations
- Community → Whole-brain entailment: community representations are contained within whole-brain representation
- These constraints force the model to learn the hierarchical organization
5. Fusion and Learning Loss
- Fuse multi-level representations for final classification
- Task-specific loss for disorder diagnosis
Experiments
- ABIDE-I: autism spectrum disorder (ASD) dataset
- REST-MDD: major depressive disorder (MDD) dataset
- HLBG outperforms state-of-the-art methods on both datasets
- Identifies disorder-relevant functional biomarkers
Key Innovations
- Hierarchical hyperbolic modeling: first to exploit hyperbolic geometry for multi-level brain graph learning
- GaMamba: novel graph-aware Mamba module with structural prompts
- Geometric entailment: formal constraints encoding brain hierarchy
- Biomarker detection: interpretable identification of disorder-relevant regions
Relationship to Existing Skills
- Extends
hyperbolic-gcn-brain-network (which uses Lorentz model for single-level brain networks)
- Adds: multi-level hierarchy, GaMamba module, geometric entailment constraints
- Complementary: HLBG is for diagnosis/classification;
hyperbolic-gcn-brain-network is for general brain network analysis
When to Use
- Brain disorder diagnosis (ASD, MDD, etc.)
- Brain network classification with hierarchical structure
- Biomarker identification from fMRI functional connectivity
- Hyperbolic graph representation learning
- When ROI-community-whole-brain hierarchy matters for the task
Pitfalls
- Lorentz model numerics: hyperbolic space operations require careful numerical handling (clipping, stability)
- Entailment constraint tuning: geometric entailment parameters need careful calibration
- Multi-level graph construction: community detection quality affects downstream performance
- GaMamba complexity: state-space model for graphs is more complex than standard GNN layers
- Dataset requirements: benefits most from datasets with clear hierarchical brain organization
