Research map / Graph learning / GNN architectures
Implicit vs Unfolded Graph Neural Networks
Yongyi Yang, Tang Liu, Yangkun Wang, Zengfeng Huang, David Wipf
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add concave rho (robust edge penalty) (ablation) | UGNN · Heterophily ASD_inf (avg) | accuracy: 71.25 -> 73.86 | Improved |
| add concave rho (robust edge penalty) (ablation) | UGNN · Heterophily ASD_sup (avg) | accuracy: 77.32 -> 76.91 | No clear effect |
| add concave rho (robust edge penalty) (ablation) | UGNN · Mettack 20% (Cora/Citeseer avg) | accuracy: 63.36 -> 70.43 | Improved |
| add concave rho (robust edge penalty) (ablation) | UGNN · Heterophily ASD_sup Minesweeper | accuracy: 91.41 -> 87.58 | Hurt |
| use UGNN with rho instead of IGNN | UGNN · Heterophily ASD_inf (avg) | accuracy: 68.45 -> 73.86 | Improved |
| use UGNN with rho instead of IGNN | UGNN · Mettack 20% (Cora/Citeseer avg) | accuracy: 63.09 -> 70.43 | Improved |
| match generating and recovery model type (UGNN to IGNN labels) | UGNN recovery model · Quasi-synthetic labels, IGNN-generated | accuracy: 95.7 -> 89.8 | Hurt |
| match generating and recovery model type (IGNN to UGNN labels) | IGNN recovery model · Quasi-synthetic labels, UGNN-generated | accuracy: 94.5 -> 91.0 | Hurt |
| restrict to truncated propagation steps (hybrid vs IGNN) on IGNN labels (ablation) | Hybrid model · Quasi-synthetic labels, IGNN-generated | accuracy: 95.7 -> 91.6 | Hurt |
| use hybrid (asymmetric, finite steps) instead of IGNN on UGNN labels | Hybrid model · Quasi-synthetic labels, UGNN-generated | accuracy: 91.0 -> 92.3 | Improved |
| use IGNN/UGNN long-range propagation instead of MLP | IGNN and UGNN · Heterophily ASD_sup (avg) | accuracy: 61.19 -> 77.32 | Improved |
| use IGNN/EIGNN/UGNN(w/o rho) long-range propagation instead of ignoring graph (MLP) | IGNN, EIGNN, UGNN w/o rho · Heterophily ASD_inf (avg) | accuracy: 71.01 -> 71.25 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- EIGNN — EIGNN: Efficient Infinite-Depth Graph Neural Networks (NeurIPS 2021) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Implicit Graph Neural Networks (NeurIPS 2020 · similarity 0.91)
- Domain Adaptive Unfolded Graph Neural Networks (AAAI 2025 · similarity 0.90)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.89)
- Classic GNNs are Strong Baselines: Reassessing GNNs for Node Classification (NeurIPS 2024 · similarity 0.89)
- MuseGNN: Forming Scalable, Convergent GNN Layers that Minimize a Sampling-Based Energy (ICLR 2025 · similarity 0.89)
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · similarity 0.88)
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.
