Research map / Continual learning / Domain and prompt-based
GCAL: Adapting Graph Models to Evolving Domain Shifts
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 GCAL, an unsupervised graph continual domain adaptation method that pairs information-maximization adaptation with replay of variational, information-bottleneck-derived memory graphs, evaluated on four evolving OOD graph datasets.
Problem
Graph domain adaptation is usually single-step and suffers catastrophic forgetting when graphs arrive sequentially under continual distribution shifts, and existing replay methods need labels to select memories.
Approach
Bilevel "adapt and generate memory" optimization: adapt the model with information maximization on new and replayed memory graphs, while a variational generator (with top-k node selection, Gumbel-style edge sampling, and a theory-derived IB lower bound) condenses each target graph into a small unlabeled memory graph via gradient matching, regularization and distribution-alignment losses.
Key results
- GCAL reaches AP-AUC 55.65% on Twitch-Explicit, versus 53.94% for the best baseline CoTTA, and AP-F1 56.57% on Elliptic versus 54.67% for Norm.
- On OGB-Arxiv GCAL achieves AP-ACC 45.22% versus 42.62% for DANN, with positive average forgetting (0.76%).
- Single-step methods such as EERM and Tent show declining performance as domains accumulate, with EERM and GTrans unable to report forgetting; GTrans runs out of memory on Facebook-100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GCAL instead of CoTTA | continual graph adaptation · Twitch-Explicit | AP-AUC: 53.94$\pm$0.36 -> 55.65$\pm$0.09 | Helped |
| use GCAL instead of DANN | continual graph adaptation · Facebook-100 | AP-ACC: 50.63$\pm$0.64 -> 52.72$\pm$0.36 | Helped |
| use GCAL instead of Norm | continual graph adaptation · Elliptic | AP-F1: 54.67$\pm$0.00 -> 56.57$\pm$0.14 | Helped |
| use GCAL instead of DANN | continual graph adaptation · OGB-Arxiv | AP-ACC: 42.62$\pm$0.23 -> 45.22$\pm$0.17 | Helped |
| add L_Reg and L_Gen (ablation) | GCAL · Twitch-Explicit, Facebook-100, Elliptic, and OGB-Arxiv | AP: 54.03$\pm$2.63 -> 55.65$\pm$0.09; 52.05$\pm$0.31 -> 52.72$\pm$0.36; 46.53$\pm$0.01 -> 56.57$\pm$0.14; 44.70$\pm$0.06 -> | Helped |
| add L_Reg (ablation) | GCAL · Twitch-Explicit, Facebook-100, Elliptic, and OGB-Arxiv | AP: 55.34$\pm$0.41 -> 55.65$\pm$0.09; 52.37$\pm$0.56 -> 52.72$\pm$0.36; 55.23$\pm$0.32 -> 56.57$\pm$0.14; 44.76$\pm$0.48 -> | Helped |
| add L_Gen (ablation) | GCAL · Twitch-Explicit, Facebook-100, Elliptic, and OGB-Arxiv | AP: 55.37$\pm$0.33 -> 55.65$\pm$0.09; 52.14$\pm$0.32 -> 52.72$\pm$0.36; 55.64$\pm$0.57 -> 56.57$\pm$0.14; 44.91$\pm$0.11 -> | Helped |
| add EMA (ablation) | GCAL · Twitch-Explicit, Facebook-100, Elliptic, and OGB-Arxiv | AP: 54.79$\pm$0.04 -> 55.65$\pm$0.09; 47.66$\pm$0.06 -> 52.72$\pm$0.36; 53.83$\pm$0.20 -> 56.57$\pm$0.14; 43.19$\pm$0.08 -> | Helped |
Surprising findings
- Removing the generation loss and regularization loss (w/o both) leaves Twitch-Explicit almost unchanged (54.03 vs 55.65 full) yet the paper attributes remaining gains to EMA and regularization, suggesting several components contribute unevenly.
- Several baselines show negative average forgetting, e.g., EATA and Norm, indicating they degrade on earlier domains after adaptation.
Limitations
- The authors note GCAL does not improve the graph model architecture itself, which may hinder performance when the base model is weak.
- Experiments use only GCN-style backbones (GCN, GraphSAGE) for main results; generalization to other architectures is shown only in an appendix analysis.
- Baselines not designed for graphs were adapted to GCNs by the authors, which may affect fairness of comparison.
Experimental setup
Datasets: Facebook-100 Twitch-Explicit OGB-Arxiv Elliptic
Benchmarks: Twitch-Explicit Facebook-100 Elliptic OGB-Arxiv
Metrics: AP (average performance) AF (average forgetting) ROC-AUC accuracy F1
Baselines: Test DANN Tent BN Stats Adapt EERM GTRANS CoTTA EATA
Builds on: information maximization (SHOT-style) graph information bottleneck gradient matching graph condensation (GCond) CoTTA EERM
Closest papers by meaning
- Efficient Memory Crystallization for Graph Learning under Non-Stationary Distribution Shifts (NeurIPS 2026 · similarity 0.90)
- Towards Continuous Reuse of Graph Models via Holistic Memory Diversification (ICLR 2025 · similarity 0.87)
- GraphKeeper: Graph Domain-Incremental Learning via Knowledge Disentanglement and Preservation (NeurIPS 2025 · similarity 0.87)
- What Matters in Graph Class Incremental Learning? An Information Preservation Perspective (NeurIPS 2024 · similarity 0.86)
- A Selective Learning Method for Temporal Graph Continual Learning (ICML 2025 · similarity 0.86)
- Exploring Rationale Learning for Continual Graph Learning (AAAI 2025 · similarity 0.86)
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.