| name | hyperbolic-learning-brain-graphs |
| description | Hyperbolic Learning on Brain Graphs (HLBG) framework for brain disorder diagnosis. Uses Lorentzian hyperbolic space to model hierarchical relationships among ROIs, functional communities, and whole-brain network. Introduces Graph-aware Mamba (GaMamba) for capturing long-range dependencies while preserving graph topology. SOTA on ABIDE-I (autism) and REST-MDD (depression) datasets. Use when: brain network analysis for disorder diagnosis, hyperbolic graph learning, brain functional connectivity modeling, hierarchical brain network representation, Graph Mamba applications, biomarker identification from fMRI, autism/depression classification from brain graphs. Trigger words: hyperbolic brain graphs, HLBG, GaMamba, Graph-aware Mamba, brain disorder diagnosis, ABIDE, REST-MDD, hierarchical brain network, Lorentzian space, functional connectivity, brain graph classification. |
Hyperbolic Learning on Brain Graphs (HLBG)
Paper
Title: Navigating Hierarchy: Hyperbolic Learning on Brain Graphs for Disorder Diagnosis
arXiv: 2607.07077v1 (2026-07-08)
Authors: Yapeng Li, Bo Jiang, Ziyan Zhang, Dongdong Chen, Zhengzheng Tu
Categories: cs.CV, cs.AI
Problem
Functional brain networks exhibit hierarchical organization across ROI, community, and whole-brain levels. Existing methods struggle to:
- Model ROI→community interactions adequately
- Capture long-range dependencies between spatially distant ROIs
- Exploit the intrinsic hierarchical geometry of brain networks
Solution: HLBG Framework
Architecture Overview
Global Brain Graph (G^g)
│
├── Hierarchical Brain Graph Construction
│ ├── Global graph: N ROIs, FC matrix, k-strongest connections
│ └── C community-specific subgraphs (Yeo 7-network parcellation)
│
├── Brain Graph Embedding via GaMamba
│ ├── Global branch: captures long-range interactions across all nodes
│ └── C Local branches: extract intra-community features
│ └── GaMamba: GAT-derived structural prompts → Mamba output matrix
│
├── Hierarchical Brain Representation Learning (HBRL)
│ ├── Project ROI/community/whole-brain reps into Lorentzian hyperbolic space
│ └── Two entailment regularization losses constrain hierarchy
│
└── Adaptive Fusion → Classification + Biomarker Detection
└── Self-attention fuses local + global representations
Key Components
1. Hierarchical Brain Graph Construction
- Global graph:
G^g = (X, A) where X from FC matrix rows, A retains k-strongest connections per node
- Local subgraphs: Partition by Yeo 7-network functional communities, extract subgraph
(X^i, A^i) for each community
- Hierarchical structure:
H = {G^g, G^1, ..., G^C}
2. Graph-aware Mamba (GaMamba)
Core innovation: inject graph structure as structural prompt into Mamba's input-dependent readout matrix.
S_i = Norm(GAT({u_j}_{j∈N(i)})) # Structural prompt from GAT
C' = W_c · (C + S) # Structure-aware output matrix
y_i = C' x_i + D u_i # Final output
- Uses GAT to derive normalized structural prompts from node neighborhoods
- Prompts integrated into Mamba's output matrix C via linear transformation
- Enables simultaneous modeling of sequential dynamics and graph-aware representations
3. Hierarchical Brain Representation Learning (HBRL)
Projects representations into Lorentzian hyperbolic space and imposes two geometric entailment constraints:
- ROI → Community entailment: ROI representations must be entailed by their community representations
- Community → Whole-brain entailment: Community representations must be entailed by global representations
This captures the one-to-many hierarchical relationships naturally in hyperbolic geometry.
Results
| Dataset | Task | Performance |
|---|
| ABIDE-I | ASD classification | SOTA |
| REST-MDD | MDD classification | SOTA |
- Identifies disorder-relevant functional biomarkers
- Outperforms GNN-based methods (BrainGB, IBGNN, BrainGNN)
- Outperforms Transformer-based methods (BrainNetTF, ALTER, Com-BrainTF, CAGT)
Implementation Patterns
GaMamba Module
class GaMamba(nn.Module):
def __init__(self, d_model, num_layers):
self.gat = GATConv(d_model, d_model)
self.mamba = Mamba(d_model=d_model)
def forward(self, X, A):
S = norm(self.gat(X, A))
return self.mamba(X, C_prime)
Hierarchical Entailment Loss
def entailment_loss(x_child, x_parent):
"""Lorentzian entailment regularization in hyperbolic space."""
r_child = lorentz_norm(x_child)
r_parent = lorentz_norm(x_parent)
return max(0, r_parent - r_child + margin)
Full Pipeline
def HLBG(global_graph, community_subgraphs):
H_global = GaMamba(global_graph.X, global_graph.A)
H_locals = [GaMamba(G.X, G.A) for G in community_subgraphs]
h_global = exp_map(H_global)
h_locals = [exp_map(H) for H in H_locals]
loss_entail = sum(entailment_loss(h_roi, h_community)
for roi, community in hierarchy_pairs)
loss_entail += entailment_loss(h_community, h_global)
fused = self_attention([h_global] + h_locals)
return classifier(fused)
Why Hyperbolic Space?
- Natural hierarchy encoding: Hyperbolic space has exponential volume growth, matching tree-like hierarchical structures
- Lorentzian model: Mathematically convenient for optimization with gradient-based methods
- Entailment geometry: Parent-child relationships in hierarchy map naturally to radial ordering in hyperbolic space
Activation
Keywords: hyperbolic learning, brain graphs, HLBG, GaMamba, Graph-aware Mamba, Lorentzian space, brain disorder diagnosis, ABIDE-I, REST-MDD, autism classification, depression diagnosis, functional connectivity, hierarchical brain network, biomarker identification, Graph Mamba, entailment regularization
References
- arXiv: 2607.07077v1
- Full paper text:
/tmp/paper_2607.07077.txt