Research map / Graph learning / GNN architectures
Deep Graph Mating
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.
Introduces Deep Graph Mating (GRAMA), a training-free, label-free way to merge pre-trained GNN parents into a child GNN via permutation-aligned parameter interpolation with topology-aware alignment (PMC) and learning-free message normalisation (CMC).
Problem
Reusing multiple pre-trained GNNs trained on disjoint data into one child GNN without re-training, fine-tuning, or labels; naive weight interpolation and neuron-alignment merging fail on graphs.
Approach
Extends weight-space model merging to GNNs by aligning neurons using aggregated graph messages (Parent Message Coordination) and then calibrating message mean/variance with parent statistics via a learning-free normalisation layer (Child Message Calibration) to counter over-smoothing.
Key results
- On ogbn-arxiv/ogbn-products node classification, the training-free DuMCC (w/ CMC) reaches 0.6645/0.6382 and 0.7647/0.7515 on the two disjoint splits, versus 0.7150/0.6687 and 0.7973/0.7775 for re-trained KA.
- On ModelNet40 point cloud classification with DGCNN, DuMCC (w/ CMC) achieves 0.8920 and 0.8574, outperforming VAPI (0.6538/0.5482) and VPI (0.4518/0.4096) and close to KA (0.9250/0.9283).
- On S3DIS semantic segmentation, DuMCC (w/ CMC) gets 0.7676 overall mIoU-style score vs 0.5060 for VAPI and 0.4791 for VPI, but below KA's 0.8382.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CMC to PMC (ablation) | DuMCC with GCN · ogbn-arxiv Dataset A | score: 0.6531 -> 0.6645 | Helped |
| Add CMC to PMC (ablation) | DuMCC with GCN · ogbn-arxiv Dataset B | score: 0.5957 -> 0.6382 | Helped |
| Add CMC to PMC (ablation) | DuMCC with GraphSAGE · ogbn-products Dataset C | score: 0.7374 -> 0.7647 | Helped |
| Add CMC to PMC (ablation) | DuMCC with GraphSAGE · ogbn-products Dataset D | score: 0.7414 -> 0.7515 | Helped |
| Add CMC to PMC (ablation) | DuMCC with DGCNN · ModelNet40 Dataset I | score: 0.8326 -> 0.8920 | Helped |
| Add CMC to PMC (ablation) | DuMCC with DGCNN · ModelNet40 Dataset J | score: 0.8088 -> 0.8574 | Helped |
| Add CMC to PMC (ablation) | DuMCC with DGCNN · S3DIS Overall | score: 0.7497 -> 0.7676 | Helped |
| Use topology-independent parameter averaging (VPI) instead of topology-aware message coordination | GRAMA · ogbn-arxiv and ogbn-products | score | Hurt |
Surprising findings
- Child GNN derived from parent message coordination shows reduced Dirichlet energy, i.e., more over-smoothing than parents, motivating the calibration step.
- Training-free DuMCC slightly outperforms the re-trained KA on multi-label molecule property prediction (ogbg-molbbbp, ogbg-molbace) and ogbn-proteins-type tasks.
Limitations
- Only homogeneous GRAMA with identical parent architectures is supported; cross-architecture and cross-level tasks are not handled.
- Performance still lags behind re-trained knowledge amalgamation on several node-level benchmarks.
- Relies on unlabelled graph data passed through parents to compute alignment statistics; fully data-independent scheme is future work.
Experimental setup
Datasets: ogbn-arxiv ogbn-products ModelNet40 S3DIS ogbn-proteins ogbg-molbace ogbg-molbbbp
Benchmarks: ogbn-arxiv ogbn-products ModelNet40 S3DIS ogbn-proteins ogbg-molbace ogbg-molbbbp
Metrics: accuracy ROC-AUC mean IoU
Baselines: KA VPI VAPI Git Re-basin weight matching
Builds on: Git Re-basin OTFusion ZipIt! REPAIR Knowledge Amalgamation
Closest papers by meaning
- A Flexible, Equivariant Framework for Subgraph GNNs via Graph Products and Graph Coarsening (NeurIPS 2024 · similarity 0.86)
- A Unified Graph Language Model for Multi-Domain Multi-Task Graph Alignment Instruction Tuning (NeurIPS 2026 · similarity 0.85)
- NN-Former: Rethinking Graph Structure in Neural Architecture Representation (CVPR 2025 · similarity 0.85)
- GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks (NeurIPS 2025 · similarity 0.85)
- Accurate and Scalable Graph Neural Networks via Message Invariance (ICLR 2025 · similarity 0.85)
- AdapterGNN: Parameter-Efficient Fine-Tuning Improves Generalization in GNNs (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.