Research map / Continual learning / Online and replay-based
Exploring Rationale Learning for Continual Graph Learning
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 RL-GNN, a rehearsal-based graph-level continual graph learning paradigm that uses invariant rationale learning with virtual environments, alternating cross-entropy and supervised contrastive training, and class prototypes at inference.
Problem
Catastrophic forgetting in graph-level continual graph learning, where models may learn spurious shortcut correlations between labels and environment subgraphs that worsen forgetting when new classes arrive.
Approach
Applies invariant learning to continual graph learning by learning a node-mask rationale generator that separates rationale and environment subgraphs, builds virtual environments by recombining latent embeddings within a mini-batch, and trains with a supervised contrastive invariant objective alternated with cross-entropy, using nearest-class-mean prototypes from a memory buffer for inference.
Key results
- On Aromaticity-CL, RL-GNN (nc=100) reaches 68.24% AP and -4.95% AF, versus 37.54% AP and 14.00% AF for ER-GNN.
- On REDDIT-MULTI-12K-CL, RL-GNN (nc=100) achieves 29.73% AP and -10.57% AF, compared with 22.60% AP for ER-GNN.
- On ENZYMES-CL, RL-GNN (nc=100) achieves 30.50% AP, above ER-GNN's 23.17% and ER-GS-LS's 25.50%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RL-GNN instead of ER-GS-LS | Graph-level continual learning · Aromaticity-CL | AP: 45.40±1.50 -> 68.24±5.68 | Helped |
| Use RL-GNN instead of ER-GS-LS | Graph-level continual learning · Aromaticity-CL | AF: 17.60±1.20 -> -4.95±3.31 | Hurt |
| Use RL-GNN instead of ER-GNN | Graph-level continual learning · REDDIT-MULTI-12K-CL | AP: 22.60±4.66 -> 29.73±2.00 | Helped |
| Use RL-GNN instead of ER-GNN | Graph-level continual learning · REDDIT-MULTI-12K-CL | AF: -32.06±4.36 -> -10.57±3.90 | Helped |
| Use RL-GNN instead of ER-GS-LS | Graph-level continual learning · ENZYMES-CL | AP: 25.50±1.70 -> 30.50±5.77 | Helped |
| Use RL-GNN instead of ER-GS-LS | Graph-level continual learning · ENZYMES-CL | AF: -20.00±5.10 -> -26.25±8.96 | Hurt |
| Increase memory-buffer allocation from nc=10 to nc=100 (ablation) | RL-GNN · Aromaticity-CL | AP: 56.00±7.85 -> 68.24±5.68 | Helped |
| Increase memory-buffer allocation from nc=10 to nc=100 (ablation) | RL-GNN · REDDIT-MULTI-12K-CL | AP: 23.04±3.79 -> 29.73±2.00 | Helped |
| Increase memory-buffer allocation from nc=10 to nc=100 (ablation) | RL-GNN · ENZYMES-CL | AP: 26.17±5.43 -> 30.50±5.77 | Helped |
Surprising findings
- Standard fine-tuning with a cross-entropy objective and naive training show strong recency bias, and the authors argue cross-entropy induces recency effects in CGL.
- Increasing memory size per class (nc) improves results, indicating dependence on representative memory samples.
Limitations
- Relies on memory buffer samples; performance depends on buffer size and representativeness.
- Sensitivity to gamma varies across datasets, requiring dataset-specific regularization tuning.
Experimental setup
Datasets: Aromaticity REDDIT-MULTI-12K ENZYMES
Benchmarks: Aromaticity-CL REDDIT-MULTI-12K-CL ENZYMES-CL
Metrics: average performance (AP) average forgetting (AF)
Baselines: Joint-train Fine-tune GEM EWC LWF MAS TWP ER-GNN ER-GS-LS
Builds on: CGLB benchmark ER-GNN invariant rationalization (Li et al. 2022) EWC nearest class mean classifier
Closest papers by meaning
- A Selective Learning Method for Temporal Graph Continual Learning (ICML 2025 · similarity 0.89)
- A Topology-aware Graph Coarsening Framework for Continual Graph Learning (NeurIPS 2024 · similarity 0.88)
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · similarity 0.88)
- Disentangled Continual Graph Neural Architecture Search with Invariant Modular Supernet (ICML 2024 · similarity 0.87)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.87)
- Dynamic Sub-graph Distillation for Robust Semi-supervised Continual Learning (AAAI 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.