Research map / Graph learning / GNN architectures
Graph Neural Networks are Inherently Good Generalizers: Insights by Bridging GNNs and MLPs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 untrained message-passing layers at inference to an MLP | MLP (PMLP) · Cora, Citeseer, Pubmed, Amazon-Photo, Amazon-Computer, Coauthor-CS, Coauthor-Physics | accuracy: +37.18%, +21.00%, +7.49%, +17.84%, +23.75%, +4.86%, +7.36% | Helped |
| Use PMLP instead of its GNN counterpart | PMLP · inductive node classification benchmarks | accuracy: +1.39%, +0.59%, -0.65%, -0.66%, -0.94%, -0.03%, +0.14% | Mixed |
| Train using MLP and apply the matching graph propagation architecture at inference | PMLP · OGBN-Arxiv | accuracy: 53.86 -> 63.74 | Helped |
| Train using MLP and apply the matching graph propagation architecture at inference | PMLP · OGBN-Products | accuracy: 56.24 -> 70.09 | Helped |
| Train using MLP and apply the matching graph propagation architecture at inference | PMLP · Flickr | accuracy: 46.44 -> 50.99; 46.44 -> 44.31 | Mixed |
| Use MLP training in place of GNN training, retaining propagation at inference | PMLP_GCN · OGBN-Arxiv | training time: 63.48 ms -> 7.78 ms | Helped |
| Use MLP training in place of GNN training, retaining propagation at inference | PMLP_GCN · OGBN-Products | training time: 288.61 ms -> 39.73 ms | Helped |
| Remove self-loops from the message-passing transition matrix (ablation) | PMLP and GNN · Cora | accuracy | Mixed |
| Replace symmetric propagation with random-walk or heat-kernel diffusion propagation (ablation) | PMLP and GNN · Cora | accuracy | No clear effect |
| Increase the number of model layers (ablation) | GCN, PMLP_GCN, and MLP · node classification benchmarks | accuracy | Hurt |
| Change the FF activation among tanh, cos, ELU, and ReLU (ablation) | GCN, PMLP, and MLP · Cora | accuracy | No clear effect |
| Reduce training proportion and available training edges (ablation) | GCN, PMLP, and MLP · Cora | accuracy | No clear effect |
Closest papers by meaning
- Towards Bridging Generalization and Expressivity of Graph Neural Networks (ICLR 2025 · similarity 0.93)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.92)
- Classic GNNs are Strong Baselines: Reassessing GNNs for Node Classification (NeurIPS 2024 · similarity 0.91)
- Covered Forest: Fine-grained generalization analysis of graph neural networks (ICML 2025 · similarity 0.91)
- A Manifold Perspective on the Statistical Generalization of Graph Neural Networks (ICML 2025 · similarity 0.90)
- How Graph Neural Networks Learn: Lessons from Training Dynamics (ICML 2024 · similarity 0.90)
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.