Research map / Graph learning / Graph self-supervised learning
What Matters in Graph Class Incremental Learning? An Information Preservation Perspective
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 GSIP, a graph spatial information preservation framework for graph class incremental learning that uses low-frequency local/global and high-frequency neighbor-similarity losses to calibrate semantic and structural shift, and plugs into replay methods (ERGNN, SSM, CaT).
Problem
Graph class incremental learning suffers catastrophic forgetting; existing methods lack a clear theoretical understanding of what information to preserve from old models, and node semantic and graph structure shifts degrade performance on old classes.
Approach
Derive from a mutual-information bound that preserving old graph information reduces forgetting, then decompose it via graph spatial filters into low-frequency local-global terms (old node outputs aligned with new node+neighbor outputs and pooled global outputs) and a high-frequency term (new neighbor-similarity distributions match old ones), added to replay methods.
Key results
- ERGNN-GSIP improves AP and AF by about 10% on Reddit relative to ERGNN.
- On Cora, ERGNN-GSIP reaches AP 71.29 and AF -36.95 vs ERGNN 65.48 and -46.09; on Citeseer AP rises from 47.65 to 61.29.
- CaT-GSIP on Arxiv (equally split into 2 classes per task) reaches AP 81.10 and AF 2.68, surpassing Joint on that setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GSIP to ERGNN | ERGNN · Reddit equally (2) | AF: -16.21 -> -5.50 | Helped |
| Add GSIP to SSM | SSM · CoraFull equally (10) | AF: -6.02 -> 0.08 | Helped |
| Add GSIP to CaT | CaT · Arxiv equally (2) | AF: -12.73 -> 3.49 | Helped |
| Add low-frequency local module (LL) (ablation) | ERGNN · CoraFull unequally | AP: 60.91 -> 65.79 | Helped |
| Add low-frequency global module (LG) after LL (ablation) | ERGNN with LL · Reddit unequally | AP: 84.63 -> 89.21 | Helped |
| Add high-frequency module (H) after LL and LG (ablation) | ERGNN with LL and LG · Reddit equally (10) | AF: -9.45 -> -2.03 | Helped |
| Add mid-frequency and/or high-frequency global information preservation (ablation) | GSIP · CoraFull equally (2) | AP: 44.79 -> 24.66 | Hurt |
Surprising findings
- CaT-GSIP on Arxiv with 2-class tasks obtains positive AF (2.68), above the Joint upper bound's implied level, as reported by the authors.
- Adding mid-frequency or high-frequency global preservation terms gives little or no further gain and can hurt (e.g., ERGNN CoraFull 2-class AF drops from -44.60 to -68.48).
Limitations
- Mid-frequency and high-frequency global preservation terms do not improve performance and add O(|M|^2) cost.
- Performance of ERGNN-GSIP is sensitive to the loss weight alpha_gip, attributed to representative-node selection causing class imbalance and topology loss.
Experimental setup
Datasets: CoraFull Arxiv Reddit Cora Citeseer
Benchmarks: CoraFull Arxiv Reddit Cora Citeseer
Metrics: Average Performance (AP) Average Forgetting (AF)
Baselines: Finetuning Joint EWC GEM MAS LwF TWP SSRM ERGNN SSM CaT
Builds on: ERGNN SSM CaT LwF TWP
Closest papers by meaning
- GraphKeeper: Graph Domain-Incremental Learning via Knowledge Disentanglement and Preservation (NeurIPS 2025 · similarity 0.89)
- Replay-and-Forget-Free Graph Class-Incremental Learning: A Task Profiling and Prompting Approach (NeurIPS 2024 · similarity 0.89)
- A Selective Learning Method for Temporal Graph Continual Learning (ICML 2025 · similarity 0.88)
- Towards Continuous Reuse of Graph Models via Holistic Memory Diversification (ICLR 2025 · similarity 0.88)
- An Efficient Memory Module for Graph Few-Shot Class-Incremental Learning (NeurIPS 2024 · similarity 0.87)
- A Topology-aware Graph Coarsening Framework for Continual Graph Learning (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.