Research map / Graph learning / Graph self-supervised learning
Information Bottleneck-Guided Adaptive Hypergraph Transformer for Brain Disease Diagnosis
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 IBAHGT, a hypergraph Transformer for brain network disease diagnosis that uses information bottleneck objectives to adaptively weight hypergraph message passing and node-level fusion of high-order and long-range features.
Problem
Brain network analysis for disease diagnosis lacks a unified way to capture high-order correlations and short/long-range dependencies, and the resulting representations contain substantial redundancy.
Approach
Combines an adaptive hypergraph convolution (KNN hypergraph, learnable node-hyperedge attention) with a Transformer encoder in parallel, and applies variational information bottleneck losses (hypergraph IB at message passing, IB at node-level fusion) with learnable per-node fusion weights to remove redundancy.
Key results
- On ABIDE (ASD vs. NC), IBAHGT reaches 77.6% ACC and 80.5% AUC, vs. 74.4% ACC for the next-best ALTER.
- On ADNI (AD vs. NC), IBAHGT reaches 86.3% ACC and 82.8% AUC, vs. 81.5% ACC for ALTER.
- Ablation on ABIDE: replacing IBAHGConv with HGNN+ drops ACC by 3.5 points and removing HIB drops ACC by 2.5 points.
Limitations
- Only functional brain networks from fMRI are studied; structural (DTI) networks are not considered.
- Reported gains on ADNI have large standard deviations (e.g., SEN 81.5±7.9), and the ADNI cohort is small (66 AD, 139 NC).
- Tuning of gamma hyperparameters is shown only for a limited grid; ADNI subset is small.
Experimental setup
Datasets: ABIDE ADNI ADHD-200
Benchmarks: ABIDE (ASD vs. NC) ADNI (AD vs. NC) ADHD-200 (ADHD vs. NC)
Metrics: ACC SPE SEN AUC
Baselines: BrainNetCNN ContrastPool BrainGNN BrainIB HGNN+ FC-HAT BrainNetTF Com-BrainTF ALTER
Builds on: information bottleneck FC-HAT BrainNetTF BrainIB HGNN+
Closest papers by meaning
- Long-range Brain Graph Transformer (NeurIPS 2024 · similarity 0.89)
- Neurodegenerative Brain Network Classification via Adaptive Diffusion with Temporal Regularization (ICML 2024 · similarity 0.85)
- Learning High-Order Relationships of Brain Regions (ICML 2024 · similarity 0.85)
- Biologically Plausible Brain Graph Transformer (ICLR 2025 · similarity 0.85)
- Causal Invariance-aware Augmentation for Brain Graph Contrastive Learning (ICML 2025 · similarity 0.84)
- PANDA: Prior-guided Attentional Dual-path Architecture (NeurIPS 2026 · 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.