Research map / Graph learning / GNN architectures
Learning from Historical Activations in Graph Neural Networks
Yaniv Galron, Hadar Sinai, Haggai Maron, Moshe Eliasof
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 signed division-by-sum normalization of layer-wise attention (vs removing it) (ablation) | HistoGraph · PROTEINS | accuracy (%): 74.45 -> 97.80 | Improved |
| add layer-wise attention over historical activations (ablation) | HistoGraph · PROTEINS | accuracy (%): 78.61 -> 97.80 | Improved |
| add node-wise self-attention at readout (ablation) | HistoGraph · PROTEINS | accuracy (%): 80.78 -> 97.80 | Improved |
| add HistoGraph readout to deep GCN | GCN · Cora, 64 layers | accuracy (%): 28.7 -> 77.5 | Improved |
| add HistoGraph readout to shallow GCN | GCN · Pubmed, 2 layers | accuracy (%): 79.0 -> 78.9 | No clear effect |
| HistoGraph vs best baseline pooling on TU with 5-layer GIN | DKEPool · RDT-B | accuracy (%): 95.0 -> 93.4 | Hurt |
| HistoGraph vs best baseline pooling on OGB molhiv | DKEPool · molhiv | ROC-AUC (%): 78.65 -> 77.81 | Hurt |
| HistoGraph vs best baseline on molbbbp | DKEPool · molbbbp | ROC-AUC (%): 69.73 -> 72.02 | Improved |
| apply HistoGraph as frozen head (FT) instead of MeanPool | pretrained GIN · IMDB-B | accuracy (%): 76.0 -> 94.0 | Improved |
| joint full fine-tuning instead of frozen head | HistoGraph post-processing · PTC | accuracy (%): 85.7 -> 97.1 | Improved |
| frozen-head fine-tune vs end-to-end HistoGraph training | HistoGraph · PROTEINS | accuracy (%): 97.3 (FT) vs 97.8 (end-to-end) | No clear effect |
| train HistoGraph end-to-end with the backbone | HistoGraph · molhiv 32-layer GCN | training time per epoch (s): 41.27 -> 60.34 | Hurt |
Closest papers by meaning
- Accurate Learning of Graph Representations with Graph Multiset Pooling (ICLR 2021 · similarity 0.87)
- Graph Multi-Convolution and Attention Pooling for Graph Classification (TPAMI 2024 · similarity 0.87)
- Haste Makes Waste: A Simple Approach for Scaling Graph Neural Networks (ICML 2025 · similarity 0.86)
- ENAHPool: The Edge-Node Attention-based Hierarchical Pooling for Graph Neural Networks (ICML 2025 · similarity 0.86)
- Geometry-Aware Edge Pooling for Graph Neural Networks (NeurIPS 2025 · similarity 0.86)
- Improving Graph Neural Networks by Learning Continuous Edge Directions (ICLR 2025 · 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.
