Research map / Graph learning / Knowledge and temporal graphs
BrainMAP: Learning Multiple Activation Pathways in Brain Networks
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 BrainMAP, which learns node orders for functional connectivity graphs with an order-learning GNN and aggregates multiple Mamba-based sequential experts via a MoE to model multiple activation pathways in fMRI brain networks.
Problem
GNNs struggle to capture long-range sequential dependencies and multiple activation pathways in fMRI functional connectivity graphs, and existing interpretability methods do not explain pathway interactions.
Approach
Adaptively sequentializes each FC graph into node orders via an order-learning GNN trained with a differentiable Spearman-style rank loss that contrasts low-loss (good) and high-loss (bad) orders, then applies a top-k attention-gated mixture of Mamba experts per order and aggregates across orders.
Key results
- BrainMAP achieves 94.74 on HCP-Task, 78.92 on HCP-Gender, and 48.44 on HCP-Age, outperforming Graph-Mamba (94.17, 77.16, 46.35) across the five HCP tasks.
- On HCP-FI and HCP-WM regression, BrainMAP reaches 10.75 and 3.81 respectively, the lowest errors among compared models.
- For MOTOR interpretation on HCP-Task, BrainMAP attains Hit@10 19.38, Hit@30 33.75, and MRR 9.26, versus 15.00, 31.25, 6.27 for Graph-Mamba.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the adaptive structure sequentializer (ablation) | BrainMAP · HCP-Task, HCP-Gender, HCP-Age, HCP-FI, HCP-WM | accuracy on classification; MAE on regression: w/o LR: 94.56, 78.09, 48.13, 3.84, 10.96; BrainMAP: 94.74, 78.92, 48.44, 3.81, 10.75 | Helped |
| add MoE pathway integration (ablation) | BrainMAP · HCP-Task, HCP-Gender, HCP-Age, HCP-FI, HCP-WM | accuracy on classification; MAE on regression: w/o MoE: 94.45, 77.58, 47.81, 3.91, 11.19; BrainMAP: 94.74, 78.92, 48.44, 3.81, 10.75 | Helped |
| add the MoE load-balancing loss (ablation) | BrainMAP · HCP-Task, HCP-Gender, HCP-Age, HCP-FI, HCP-WM | accuracy on classification; MAE on regression: w/o LB: 94.45, 78.50, 47.54, 3.91, 11.01; BrainMAP: 94.74, 78.92, 48.44, 3.81, 10.75 | Helped |
| use BrainMAP instead of the strongest reported baseline | HCP prediction benchmarks · HCP-Task, HCP-Gender, HCP-Age, HCP-FI, HCP-WM | accuracy on classification; MAE on regression: HCP-Task: 94.17~$(\pm 0.86)$ -> 94.74~$(\pm 0.07)$; HCP-Gender: 77.16~$(\pm 3.13)$ -> 78.92~$(\pm 0.49)$; HCP-Age: 46.35 | Helped |
Surprising findings
- Ablation shows removing the MoE hurts performance most, and the number of experts plateaus beyond 4.
Limitations
- Relies on the NeuroGraph benchmark and HCP datasets, limiting generality to other neuroimaging data.
- Interpretation evaluated only on the MOTOR task with ground truth from domain experts over a small set of samples.
- Higher training cost than some baselines on several datasets.
Experimental setup
Datasets: HCP-Task HCP-Gender HCP-Age HCP-FI HCP-WM
Benchmarks: HCP-Task HCP-Gender HCP-Age HCP-FI HCP-WM
Metrics: accuracy MAE Hit@10 Hit@30 MRR training time per epoch
Baselines: GCN GAT GraphSAGE ResGCN GraphGPS Graph-Mamba
Builds on: Mamba GraphGPS Graph-Mamba GNNExplainer Spearman's rank correlation
Closest papers by meaning
- BrainOOD: Out-of-distribution Generalizable Brain Network Analysis (ICLR 2025 · similarity 0.87)
- NeuroTree: Hierarchical Functional Brain Pathway Decoding for Mental Health Disorders (ICML 2025 · similarity 0.87)
- $\textit{NeuroPath}$: A Neural Pathway Transformer for Joining the Dots of Human Connectomes (NeurIPS 2024 · similarity 0.87)
- Functional Connectomes of Neural Networks (AAAI 2025 · similarity 0.86)
- Long-range Brain Graph Transformer (NeurIPS 2024 · similarity 0.85)
- Beyond Pairwise Connections: Extracting High-Order Functional Brain Network Structures under Global Constraints (NeurIPS 2025 · similarity 0.84)
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.