Research map / Graph learning / Graph self-supervised learning
Neurodegenerative Brain Network Classification via Adaptive Diffusion with Temporal Regularization
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 AGT, a wavelet-based GNN with node-wise adaptive diffusion scales and a group-level temporal regularizer, for classifying neurodegenerative stages from brain networks on ADNI and PPMI.
Problem
Cross-sectional brain network classifiers ignore the progressive, sequential nature of neurodegenerative diagnostic groups and struggle with mixed homophilic and heterophilic structure in brain graphs.
Approach
Replaces global diffusion filters with a node-variant adaptive wavelet basis whose trainable per-node (and per-class) scales can be low- or high-pass, plus a triangular-distance regularizer that aligns class-mean embeddings along the disease progression order.
Key results
- AGT reached 90.3% accuracy on ADNI-CT (5 labels), 2.5%p above the second-best method.
- AGT reached 94.8% accuracy on ADNI-FDG, 4.4%p above the second-best method.
- AGT reached 83.6% accuracy on PPMI (3 labels), 2.4%p above the second-best; temporal regularizer added about +1.0%p accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AGT instead of LSAP | ADNI-CT classification · ADNI-CT | accuracy (%): +2.5 | Helped |
| use AGT instead of LSAP | ADNI-FDG classification · ADNI-FDG | accuracy (%): +4.4 | Helped |
| use AGT instead of GAT | PPMI classification · PPMI | accuracy (%): +2.4 | Helped |
| add temporal regularization Rtemp (ablation) | AGT · ADNI-CT | accuracy (%): +0.9 | Helped |
| add temporal regularization Rtemp (ablation) | AGT · ADNI-FDG | accuracy (%): +0.4 | Helped |
| add temporal regularization Rtemp (ablation) | AGT · PPMI | accuracy (%): +1.0 | Helped |
| replace adaptive local filtering with fixed high-pass scale -2 (ablation) | AGT · ADNI-CT | accuracy (%) | Hurt |
| replace adaptive local filtering with fixed high-pass scale -1 (ablation) | AGT · ADNI-CT | accuracy (%) | Hurt |
| replace adaptive local filtering with fixed high-pass scale -0.5 (ablation) | AGT · ADNI-CT | accuracy (%) | Hurt |
| replace adaptive local filtering with fixed low-pass scale 0.5 (ablation) | AGT · ADNI-CT | accuracy (%) | Hurt |
| replace adaptive local filtering with fixed low-pass scale 1 (ablation) | AGT · ADNI-CT | accuracy (%) | Hurt |
| replace adaptive local filtering with fixed low-pass scale 2 (ablation) | AGT · ADNI-CT | accuracy (%) | Hurt |
Surprising findings
- Fixed high-pass or low-pass scales drop accuracy to 72-75% on ADNI-CT, far below adaptive scales, indicating node-wise scale learning matters.
Limitations
- Gains from the temporal regularizer are modest (about 1 percentage point).
- PPMI has only 195 subjects and high variance across folds.
- Class-wise scales rely on pseudo-labels from a pretrained Exact model.
Experimental setup
Datasets: ADNI PPMI
Benchmarks: ADNI PPMI
Metrics: accuracy precision recall specificity
Baselines: SVM MLP GCN GAT GDC GraphHeat ADC Exact LSAP BrainGNN BrainNetTF
Builds on: Exact Wave-GD LSAP GPR-GNN GraphHeat
Closest papers by meaning
- NeuroH-TGL: Neuro-Heterogeneity Guided Temporal Graph Learning Strategy for Brain Disease Diagnosis (NeurIPS 2025 · similarity 0.89)
- Data-Driven Network Neuroscience: On Data Collection and Benchmark (NeurIPS 2023 · similarity 0.88)
- Learning to Approximate Adaptive Kernel Convolution on Graphs (AAAI 2024 · similarity 0.88)
- Long-range Brain Graph Transformer (NeurIPS 2024 · similarity 0.86)
- Information Bottleneck-Guided Adaptive Hypergraph Transformer for Brain Disease Diagnosis (NeurIPS 2026 · similarity 0.85)
- A Generalized Neural Diffusion Framework on Graphs (AAAI 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.