Research map / Continual learning / Online and replay-based
Capacity-Agnostic Parameter Isolation for Continual Graph Learning
Ye Xiao, Ruikun Li, Zhenyu Yang, Andrey Vasnev, Junbin Gao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CAGNN decoupled subnetwork instead of WSN full-network propagation | parameter-isolation GNN · OGB-Arxiv | runtime per epoch (ms): 787 -> 197 | Improved |
| use CAGNN instead of WSN | parameter-isolation GNN · Synthetic CoraFull (140 tasks) | memory (MB) at T140: 8980.4 -> 6296.2 | Improved |
| use task-specific context instead of random/all-ones/parameterized scores (ablation) | CAGNN · Synthetic CoraFull (sparsity 0.99) | parameter usage: 61.30% -> 22.45% | Improved |
| use task-specific context instead of random/all-ones/parameterized scores (ablation) | CAGNN · Synthetic CoraFull (sparsity 0.99) | AP: 96.51±1.1 -> 97.05±0.3 | No clear effect |
| add differentiable proxy for mask learning (ablation) | CAGNN · Aromaticity (sparsity 0.90) | AP: 45.43±2.1 -> 49.28±1.2 | Improved |
| add differentiable proxy for mask learning (ablation) | CAGNN · Cora (sparsity 0.90) | AP: 93.97±1.2 -> 94.41±1.6 | No clear effect |
| add Fisher regularization (ablation) | CAGNN · CoraFull (parameter utilization) | parameter usage | Improved |
| use both dendritic and low-rank masking instead of one stage (ablation) | CAGNN · Computers (5 tasks) | parameter usage: 44.50% -> 30.20% | Improved |
| use both dendritic and low-rank masking instead of one stage (ablation) | CAGNN · MNIST (5 tasks) | AP: 96.89±1.2 -> 96.04±1.7 | No clear effect |
| use low-rank masking instead of neuron-wise masking (ablation) | CAGNN · synthetic CoraFull | convergence | Improved |
| use only 10% warm-up epochs instead of longer warm-up (ablation) | CAGNN · CoraFull | parameter utilization | No clear effect |
| increase sparsity / overlap ratio (ablation) | CAGNN · CoraFull (35 tasks) | accuracy vs zero-interference | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Piggyback — Piggyback: Adapting a Single Network to Multiple Tasks by Learning to Mask Weights (ECCV 2018) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Exploring Rationale Learning for Continual Graph Learning (AAAI 2025 · similarity 0.89)
- Learning without Isolation: Pathway Protection for Continual Learning (ICML 2025 · similarity 0.88)
- A Topology-aware Graph Coarsening Framework for Continual Graph Learning (NeurIPS 2024 · similarity 0.87)
- InTAct: Interval-based Task Activation Consolidation for Continual Learning (NeurIPS 2026 · similarity 0.87)
- Rethinking the Stability-Plasticity Trade-off in Continual Learning from an Architectural Perspective (ICML 2025 · similarity 0.87)
- Disentangled Continual Graph Neural Architecture Search with Invariant Modular Supernet (ICML 2024 · 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.
