Research map / Continual learning / Class-incremental
An Efficient Memory Module for Graph Few-Shot Class-Incremental 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 Mecoin, a memory-based graph few-shot class-incremental learning method that caches class prototypes in a Structured Memory Unit and stores class probability distributions in a Memory Representation Adaptive Module, distilling them into a frozen-encoder GNN to reduce forgetting.
Problem
Graph few-shot class-incremental learning (GFSCIL): GNNs must learn new node classes from very few labels while retaining prior knowledge, without caching large meta-learning sample sets or extensively fine-tuning parameters.
Approach
Decouple class prototype learning from node probability learning: prototypes are built via self-attention interaction with cached prototypes plus local graph-structure (GraphInfo) features and k-means, while per-class probability distributions are stored in a separate memory and distilled into the GNN/MLP with KL losses, so prototype updates do not alter stored class knowledge.
Key results
- On CoraFull, Mecoin (GCN) reports average accuracy 70.90% versus 35.52% for Geometer and 66.57% for HAG-Meta, with PD 20.82% (table-reported).
- On CS, Mecoin (GCN) reaches average accuracy 77.96% with PD 38.41%, compared with 28.11% for Geometer.
- On Computers, Mecoin (GCN) reaches average accuracy 77.39% versus 27.19% for Geometer; the paper also claims a 66.72% improvement over the best baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Mecoin instead of the strongest listed baseline | GCN-based GFSCIL · CoraFull | Average ACC: 66.57 -> 70.90 | Helped |
| Use Mecoin instead of the strongest listed baseline | GCN-based GFSCIL · Computers | Average ACC: 44.75 -> 77.39 | Helped |
| Use Mecoin instead of the strongest listed baseline | GCN-based GFSCIL · CS | Average ACC: 60.68 -> 77.96 | Helped |
| Add MeCs for prototype construction (ablation) | Mecoin · CoraFull, CS, and Computers | accuracy | Helped |
| Add node-feature interaction with class prototypes stored in SMU (ablation) | Mecoin · CoraFull, CS, and Computers | accuracy | Mixed |
| Add GraphInfo to MeCs (ablation) | Mecoin · CoraFull, CS, and Computers | forgetting rate | Helped |
| Use GraphInfo dimension configuration 13-1 and concatenate it left of node features (ablation) | Mecoin · CoraFull | Average ACC: 71.83 | Helped |
| Use GraphInfo dimension configuration 12-2 and concatenate it left of node features (ablation) | Mecoin · CoraFull | Average ACC: 70.94 | Helped |
| Use GraphInfo dimension configuration 11-3 and concatenate it left of node features (ablation) | Mecoin · CoraFull | Average ACC: 70.69 | Helped |
| Use GraphInfo dimension configuration 10-4 and concatenate it left of node features (ablation) | Mecoin · CoraFull | Average ACC: 70.15 | Helped |
| Use GraphInfo dimension configuration 9-5 and concatenate it left of node features (ablation) | Mecoin · CoraFull | Average ACC: 70.81 | Helped |
| Increase GraphInfo dimension in the tested configurations (ablation) | Mecoin · CoraFull | Average ACC | Hurt |
Surprising findings
- Baselines such as EWC, LwF, TWP and ER-GNN show identical or near-identical accuracy to Mecoin on Computers and CoraFull session 0-1, suggesting limited differentiation in some settings.
- HAG-Meta and Geometer perform poorly under the paper's no-meta-training fair-comparison setting, while they do better under their original settings.
Limitations
- Assumes all session graphs originate from the same graph; multi-source graphs not explored.
- Not validated under low-resource conditions where graph structural information is scarce.
- Several theoretical results rely on coarse bounds and cited proofs with unclear details.
Experimental setup
Datasets: CoraFull CS Computers
Benchmarks: CoraFull CS Computers
Metrics: average accuracy forgetting rate (PD) per-session accuracy
Baselines: EWC LwF TWP GEM ER-GNN MAS HAG-Meta Geometer
Builds on: Elastic Weight Consolidation Topology-aware Weight Preserving HAG-Meta Geometer
Closest papers by meaning
- What Matters in Graph Class Incremental Learning? An Information Preservation Perspective (NeurIPS 2024 · similarity 0.87)
- Towards Continuous Reuse of Graph Models via Holistic Memory Diversification (ICLR 2025 · similarity 0.87)
- Efficient Memory Crystallization for Graph Learning under Non-Stationary Distribution Shifts (NeurIPS 2026 · similarity 0.85)
- A Topology-aware Graph Coarsening Framework for Continual Graph Learning (NeurIPS 2024 · similarity 0.85)
- GraphKeeper: Graph Domain-Incremental Learning via Knowledge Disentanglement and Preservation (NeurIPS 2025 · similarity 0.85)
- A Selective Learning Method for Temporal Graph Continual Learning (ICML 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.