Research map / Continual learning / Class-incremental
Replay-and-Forget-Free Graph Class-Incremental Learning: A Task Profiling and Prompting Approach
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 TPP, a replay-free and forget-free graph class-incremental learning method that predicts task IDs via Laplacian-smoothing task prototypes and learns a small graph prompt per task on a frozen GNN.
Problem
Graph class-incremental learning (GCIL) suffers from catastrophic forgetting and inter-task class separation because task IDs are unavailable at inference time.
Approach
Profile each graph task with a Laplacian-smoothed prototype so test graphs can be assigned to their task ID (proven to converge with large smoothing steps), and learn a small task-specific graph prompt over a frozen GNN pre-trained once on the first task, so each task effectively gets its own classifier without replay.
Key results
- TPP achieves 100% task ID prediction accuracy on all four datasets and AF of 0.0 on all four datasets.
- TPP reaches AA of 93.4% (CoraFull), 85.4% (Arxiv), 99.5% (Reddit), 94.0% (Products), outperforming the best competing method by over 18% average AA per the paper.
- Adding the task ID predictor to existing TWP and DeLoMe boosts their AA substantially (e.g., TWP on Arxiv from 6.7% to 89.4%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TPP instead of DeLoMe | GCIL · CoraFull | AA: 81.0$\pm$0.2 -> 93.4$\pm$0.4 | Helped |
| use TPP instead of DeLoMe | GCIL · Arxiv | AA: 50.6$\pm$0.3 -> 85.4$\pm$0.1 | Helped |
| use TPP instead of DeLoMe | GCIL · Reddit | AA: 97.4$\pm$0.1 -> 99.5$\pm$0.0 | Helped |
| use TPP instead of CaT | GCIL · Products | AA: 70.3$\pm$0.9 -> 94.0$\pm$0.5 | Helped |
| remove task ID prediction (ablation) | TPP · CoraFull | AA: 93.4 -> 2.0 | Hurt |
| remove task ID prediction (ablation) | TPP · Arxiv | AA: 85.4 -> 3.0 | Hurt |
| remove task ID prediction (ablation) | TPP · Reddit | AA: 99.5 -> 2.8 | Hurt |
| remove task ID prediction (ablation) | TPP · Products | AA: 94.0 -> 2.7 | Hurt |
| add graph prompt (ablation) | TPP with task ID prediction and classification head · CoraFull | AA: 73.8 -> 93.4 | Helped |
| add graph prompt (ablation) | TPP with task ID prediction and classification head · Arxiv | AA: 76.3 -> 85.4 | Helped |
| add graph prompt (ablation) | TPP with task ID prediction and classification head · Reddit | AA: 98.6 -> 99.5 | Helped |
| add graph prompt (ablation) | TPP with task ID prediction and classification head · Products | AA: 90.0 -> 94.0 | Helped |
Surprising findings
- TPP exceeds the Joint training oracle by a large margin on all datasets, and is close to the Oracle Model that is given true task IDs.
- Node-attribute-only task prototypes already predict task IDs well on Arxiv and Reddit but fail on tasks with similar attributes, motivating the graph-structure smoothing.
Limitations
- Theoretical guarantee assumes graphs are not isolated; isolated nodes require a heuristic augmentation.
- Random task formulation on Products causes heavy class imbalance and degraded accuracy for both TPP and the Oracle Model; debiased learning left for future work.
- Only node classification on four benchmarks with two classes per task; inference adds a small task-ID overhead.
Experimental setup
Datasets: CoraFull Arxiv Reddit Products
Benchmarks: CoraFull Arxiv Reddit Products
Metrics: average accuracy (AA) average forgetting (AF) task ID prediction accuracy
Baselines: EWC LwF GEM MAS ERGNN TWP SSM SEM CaT DeLoMe OODCIL Fine-Tune
Builds on: graph contrastive learning CGLB graph continual learning benchmark OOD detection-based class-incremental learning
Closest papers by meaning
- What Matters in Graph Class Incremental Learning? An Information Preservation Perspective (NeurIPS 2024 · similarity 0.89)
- GACL: Exemplar-Free Generalized Analytic Continual Learning (NeurIPS 2024 · similarity 0.89)
- Class Incremental Learning via Likelihood Ratio Based Task Prediction (ICLR 2024 · similarity 0.88)
- GraphKeeper: Graph Domain-Incremental Learning via Knowledge Disentanglement and Preservation (NeurIPS 2025 · similarity 0.87)
- Towards Continuous Reuse of Graph Models via Holistic Memory Diversification (ICLR 2025 · similarity 0.87)
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (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.