Research map / Graph learning / GNN architectures
Better with Less: A Data-Active Perspective on Pre-Training Graph Neural Networks
Jiarong Xu, Renhong Huang, XIN JIANG, Yuxuan Cao, Carl Yang, Chunping Wang, Yang Yang
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 graph-property criteria to the graph selector (vs APT-G) (ablation) | APT, fine-tune · dd242 node classification | micro F1: 25.45 -> 28.62 | Improved |
| add predictive-uncertainty criterion to the selector (vs APT-P) (ablation) | APT, fine-tune · dd242 node classification | micro F1: 25.19 -> 28.62 | Improved |
| add proximal regularization w.r.t. old knowledge (vs APT-R) (ablation) | APT, fine-tune · ogbarxiv node classification | micro F1: 18.41 -> 27.40 | Improved |
| use Fisher-weighted proximal term instead of L2 (vs APT-L2) (ablation) | APT, fine-tune · node classification (10 graphs) | average rank: 2.6 -> 1.3 | Improved |
| add graph-property criteria to the selector (vs APT-G) (ablation) | APT, freeze · node classification (10 graphs) | average rank: 6.3 -> 3.8 | Improved |
| add predictive uncertainty to selector (vs APT-P) (ablation) | APT, freeze · node classification (10 graphs) | average rank: 6.2 -> 3.8 | Improved |
| add proximal regularization (vs APT-R) (ablation) | APT, freeze · node classification (10 graphs) | average rank: 5.0 -> 3.8 | Improved |
| Fisher-weighted proximal term instead of L2 (vs APT-L2) (ablation) | APT, freeze · node classification (10 graphs) | average rank: 4.4 -> 3.8 | Improved |
| select a subset of graphs and samples instead of pre-training on all data | GCC backbone, fine-tune · dd242 node classification | micro F1: 19.32 -> 28.62 | Improved |
| select a subset of graphs instead of all (average over datasets, freeze) | GCC backbone, freeze · node classification (10 graphs) | micro F1: 9.94% average gain | Improved |
| data selection vs training on all graphs, graph classification | GCC backbone, freeze · graph classification (3 datasets) | micro F1: 7.2% average gain | Improved |
| data selection vs training on all graphs, graph classification, fine-tune | GCC backbone, fine-tune · msrc-21 graph classification | micro F1: 24.90 -> 24.41 | No clear effect |
Closest papers by meaning
- Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025 · similarity 0.88)
- On the Interplay between Graph Structure and Learning Algorithms in Graph Neural Networks (ICML 2025 · similarity 0.88)
- Unlocking the Potential of Black-box Pre-trained GNNs for Graph Few-shot Learning (AAAI 2025 · similarity 0.88)
- How Graph Neural Networks Learn: Lessons from Training Dynamics (ICML 2024 · similarity 0.88)
- Edge Prompt Tuning for Graph Neural Networks (ICLR 2025 · similarity 0.88)
- DFA-GNN: Forward Learning of Graph Neural Networks by Direct Feedback Alignment (NeurIPS 2024 · similarity 0.87)
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.
