Research map / Graph learning / Graph self-supervised learning
Core Knowledge Learning Framework for Graph
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.
Proposes Core Knowledge Learning (CKL), which learns a task-relevant core subgraph via node/edge masking with Gumbel-softmax and uses it for graph domain adaptation (WL kernel label transfer) and few-shot molecular property prediction (bi-level meta-learning).
Problem
Graph classification struggles with adapting across diverse tasks, domain shift between source and target graphs, and small-sample prediction scenarios.
Approach
Learns a core subgraph that drives GNN predictions (treating the rest as task-irrelevant) via a differentiable node and edge selection module, then uses the core subgraphs for domain adaptation via WL subtree kernel matching and for few-shot learning via bi-level optimization.
Key results
- On Mutagenicity edge-density domain adaptation, CKL reaches 74.9% average accuracy versus 72.2% for the best baseline CoCo.
- In few-shot molecular property prediction, CKL improves average ROC-AUC by 1.62% over the strongest baseline EGNN.
- CKL achieves 82.3 (10-shot) and 81.4 (1-shot) ROC-AUC on Tox21, exceeding PAR and EGNN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CKL instead of CoCo for graph domain adaptation | Mutagenicity · Edge density domain shift (Avg.) | accuracy (%): 72.2 -> 74.9 | Helped |
| Use CKL instead of the strongest listed baseline | Few-shot molecular property prediction · Tox21 10-shot and 1-shot | ROC-AUC: 82.1±0.1 -> 82.3±0.3; 79.4±0.2 -> 81.4±0.4 | Helped |
| Use CKL instead of the strongest listed baseline | Few-shot molecular property prediction · SIDER 10-shot | ROC-AUC: 74.7±0.3 -> 75.3±0.5 | Helped |
| Use CKL instead of PAR | Few-shot molecular property prediction · SIDER 1-shot | ROC-AUC: 71.9±0.5 -> 71.7±0.7 | Hurt |
| Use CKL instead of the strongest listed baseline | Few-shot molecular property prediction · MUV 10-shot and 1-shot | ROC-AUC: 66.5±2.1 -> 67.1±2.3; 64.1±1.2 -> 64.4±2.2 | Helped |
| Use CKL instead of PAR | Few-shot molecular property prediction · ToxCast 10-shot and 1-shot | ROC-AUC: 69.7±1.6 -> 70.2±1.5; 67.3±2.9 -> 68.1±1.6 | Helped |
| Replace GIN with other GNN backbones or replace the WL subtree kernel with other graph kernels (ablation) | CKL graph domain adaptation · Graph datasets in Figure 5 | performance | Mixed |
Surprising findings
- GIN and WL subtree kernel were the top performers across GNN and kernel variants, suggesting the choice of backbone matters more than the framework's flexibility.
Limitations
- Relies on WL subtree kernel for source-target matching, with limited expressiveness shown only qualitatively.
- Few-shot evaluation omits Siamese and IterRefLSTM on ToxCast due to missing results.
- Claims about the core subgraph's role rest on the Gilbert random graph independence assumption, which is a simplification.
Experimental setup
Datasets: Mutagenicity Tox21 AhR FRANKENSTEIN PROTEINS DD COX2 COX2 MD BZR BZR MD Tox21 SIDER MUV
Benchmarks: Mutagenicity (edge density domain shift) Tox21 SIDER MUV ToxCast
Metrics: accuracy ROC-AUC
Baselines: WL subtree GCN GIN GMT CIN CDAN ToAlign MetaAlign DEAL CoCo Siamese ProtoNet
Builds on: GNNExplainer GIN MAML CoCo Weisfeiler-Lehman subtree kernel
Closest papers by meaning
- Graph Classification via Reference Distribution Learning: Theory and Practice (NeurIPS 2024 · similarity 0.89)
- No Metric to Rule Them All: Toward Principled Evaluations of Graph-Learning Datasets (ICML 2025 · similarity 0.88)
- Open-Set Graph Domain Adaptation via Separate Domain Alignment (AAAI 2024 · similarity 0.88)
- Hypergraph-enhanced Dual Semi-supervised Graph Classification (ICML 2024 · similarity 0.87)
- Subgraph Invariant Learning Towards Large-Scale Graph Node Classification (AAAI 2025 · similarity 0.87)
- Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation (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.