Research map / Graph learning / Knowledge and temporal graphs
Long-range Brain Graph Transformer
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 ALTER, a brain graph transformer that injects long-range embeddings from correlation-biased random walks into self-attention to diagnose neurological disease from fMRI brain graphs.
Problem
Brain network analysis with graph learning focuses on short-range neighborhood aggregation and neglects long-range inter-ROI dependencies that reflect brain-wide communication.
Approach
An adaptive long-range aware (ALGA) random-walk encoding whose transition probabilities are reweighted by inter-ROI Pearson correlations, whose resulting step-wise embeddings are injected as tokens into a vanilla transformer encoder so self-attention integrates short- and long-range dependencies.
Key results
- ALTER reaches 82.8% AUC and 77.0% ACC on ABIDE, outperforming the best specialized baseline BrainNETTF (80.2% AUC, 71.0% ACC).
- On ADNI, ALTER achieves 78.8% AUC and 74.1% ACC versus 76.5% AUC and 69.0% ACC for BrainNETTF.
- Adding ALGA to SAN and Graphormer improves ABIDE AUC from 71.3 to 72.5 and from 63.5 to 67.2, respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ALGA to ALTER (ablation) | ALTER · ABIDE | ACC: 71.0~(1.2) -> 77.0~(1.0) | Helped |
| add ALGA to ALTER (ablation) | ALTER · ADNI | ACC: 69.0~(2.7) -> 74.1~(2.5) | Helped |
| add adaptive factors to random-walk encoding at 16 hops (ablation) | ALTER · ABIDE | AUC: 81.2(2.6) -> 82.8(1.1) | Helped |
| add adaptive factors to random-walk encoding at 16 hops (ablation) | ALTER · ADNI | AUC: 78.1(2.9) -> 78.8~(2.1) | Helped |
| increase random-walk length from 2 to 16 hops (ablation) | ALTER · ABIDE | AUC: 78.7(3.9) -> 82.8(1.1) | Helped |
| increase random-walk length from 2 to 16 hops (ablation) | ALTER · ADNI | AUC: 77.1(2.9) -> 78.8~(2.1) | Helped |
| replace clustering readout with mean, max, sum, or sort pooling (ablation) | VanillaTF with ALGA on ABIDE · ABIDE | AUC: 82.8(1.1) -> 75.6(1.6); 76.4(1.7); 72.0(1.2); 74.4(1.4) | Hurt |
Surprising findings
- Graphormer and GraphTrans perform near chance on ADNI, and several baselines (e.g., BrainNetGNN, A-GCL) show very large variance in sensitivity/specificity.
Limitations
- Balance between short- and long-range dependencies is not guaranteed optimal, as the authors acknowledge.
- Only fMRI functional connectivity is used; other modalities such as DTI are not explored.
- Group-level long-range dependence is weaker than individual-level, with individual differences possibly confounding.
Experimental setup
Datasets: ABIDE ADNI
Benchmarks: ABIDE ADNI
Metrics: AUC ACC F1-Score SEN SPE
Baselines: SAN Graphormer GraphTrans LRGNN FBNETGEN BrainNetGNN BrainGNN BrainNETTF A-GCL ContrastPool
Builds on: Graphormer SAN random walk structural encoding GraphTrans LRGNN
Closest papers by meaning
- Information Bottleneck-Guided Adaptive Hypergraph Transformer for Brain Disease Diagnosis (NeurIPS 2026 · similarity 0.89)
- Biologically Plausible Brain Graph Transformer (ICLR 2025 · similarity 0.87)
- Adaptive Message Passing: A General Framework to Mitigate Oversmoothing, Oversquashing, and Underreaching (ICML 2025 · similarity 0.86)
- On Measuring Long-Range Interactions in Graph Neural Networks (ICML 2025 · similarity 0.86)
- MAGNET: Manifold-Aware Graph Diffusion Network for Connectome Generation (NeurIPS 2026 · similarity 0.86)
- Neurodegenerative Brain Network Classification via Adaptive Diffusion with Temporal Regularization (ICML 2024 · similarity 0.86)
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.