Research map / Continual learning / Domain and prompt-based

GraphKeeper: Graph Domain-Incremental Learning via Knowledge Disentanglement and Preservation

NeurIPS 2025 · Continual learning › Domain and prompt-based · approach: Domain-incremental learning with memory · 1 citations (OpenAlex)

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 GraphKeeper, a graph domain-incremental learning framework that combines domain-specific LoRA on a frozen pre-trained GNN, intra/inter-domain disentanglement losses, and recursive ridge-regression classifiers to prevent forgetting across graph domains.

Problem

Catastrophic forgetting in graph domain-incremental learning (Domain-IL), where sequentially arriving graphs come from different domains, a setting previously unexplored in graph incremental learning.

Approach

Decomposes forgetting into embedding shifts and decision-boundary deviations: freezes domain-specific LoRA modules to stabilize embeddings, adds intra- and inter-domain contrastive/scattering objectives, and replaces the end-to-end classifier with a closed-form, recursively updated ridge regression that needs no historical data; test graphs of unknown domain are routed via random-projection domain

Key results

• GraphKeeper achieves 6.5%–16.6% AA improvement over the runner-up across the six Domain-IL groups with negligible forgetting (e.g., 69.2% AA on Group 1 vs. 52.6% for TPP).
• On longer 11- and 12-domain sequences, GraphKeeper reaches 73.0% and 76.5% AA, outperforming the runner-up by 22.8%–23.7%.
• Integrating GraphKeeper with GCOPE and MDGPT in few-shot Domain-IL raises AA substantially (e.g., GCOPE from 20.6% to 56.8% on Group 1) with near-zero forgetting.

Measured effects reported in the paper

Surprising findings

• GraphKeeper outperforms the Joint baseline (training on all previous domains) on all groups, suggesting a single GNN struggles to integrate multi-domain knowledge.
• Baselines DeLoMe and PDGNNs degrade notably when their SGC/APPNP backbones are replaced by GCN.

Limitations

• Assumes classes in each new incremental domain are entirely novel; partially overlapping classes across domains are not explored.
• Does not exploit zero-shot ability of GFMs or handle previously unseen domains at test time.
• Domain discrimination for unobservable-domain test graphs relies on prototype matching in random-projection space.

Experimental setup

Datasets: Cora Citeseer PubMed CoauthorCS DBLP Photo Computer WikiCS Facebook Chameleon Squirrel GitHub

Benchmarks: Domain-IL incremental groups 1-6 (15 datasets) longer Domain-IL sequences Group 7 and Group 8 few-shot Domain-IL with GCOPE and MDGPT

Metrics: Average Accuracy (AA) Average Forgetting (AF)

Baselines: EWC MAS GEM LWF TWP ER-GNN SSM DeLoMe PDGNNs TPP

Builds on: LoRA CGLB benchmark DBSCAN

Closest papers by meaning

• What Matters in Graph Class Incremental Learning? An Information Preservation Perspective (NeurIPS 2024 · similarity 0.89)
• Towards Continual Knowledge Graph Embedding via Incremental Distillation (AAAI 2024 · similarity 0.88)
• Open-Set Graph Domain Adaptation via Separate Domain Alignment (AAAI 2024 · similarity 0.88)
• How Much Can Transfer? BRIDGE: Bounded Multi-Domain Graph Foundation Model with Generalization Guarantees (ICML 2025 · similarity 0.88)
• Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 2025 · similarity 0.88)
• Towards Unsupervised Open-Set Graph Domain Adaptation via Dual Reprogramming (NeurIPS 2025 · similarity 0.87)