Research map / Continual learning / Online and replay-based
Disentangled Continual Graph Neural Architecture Search with Invariant Modular Supernet
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 GASIM, a continual graph neural architecture search method that uses a modular supernet, a factor-based task-module router, and an invariance loss to avoid forgetting across sequential node-classification tasks.
Problem
Graph neural architecture search assumes static graph tasks; in continual settings, searching a new architecture for a new task causes architecture conflicts and catastrophic forgetting of past tasks.
Approach
Disentangle the search by maintaining K architecture modules in a supernet, routing each incoming task to one module via latent graph factors and a task-module relation graph (hard routing, only that module is searched), and adding an invariance loss over tasks sharing a module so the architecture keeps handling previously routed tasks.
Key results
- On CoraFull, GASIM reaches 76.03% AP versus 67.70% for the best manually designed baseline (GIN) and 66.02% for DARTS.
- On Arxiv, GASIM reaches 78.95% AP (best among methods; DARTS 77.40%) with AF -12.19 versus DARTS -16.59 on the mean-AP-style metric reported in Table 2.
- On Reddit, GASIM reaches 99.10% AP, the highest of all methods (GIN 98.68%, DARTS 98.75%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GASIM instead of GraphConv | Continual graph neural architecture search · CoraFull | AP: 67.15±1.12 -> 76.03±0.77 | Helped |
| Use GASIM instead of DARTS | Continual graph neural architecture search · Arxiv | AP: 61.45±2.09 -> 65.03±1.40 | Helped |
| Use GASIM instead of Random | Continual graph neural architecture search · Reddit | AP: 97.89±0.08 -> 98.40±0.01 | Helped |
| Remove the invariance loss (ablation) | GASIM · CoraFull, Arxiv, Reddit | mean AP | Hurt |
| Select random factors instead of similar factors (ablation) | GASIM · CoraFull, Arxiv, Reddit | mean AP | Hurt |
| Replace the factor-based router with a random router (ablation) | GASIM · CoraFull, Arxiv, Reddit | mean AP | Hurt |
| Replace the factor-based router with a fixed router (ablation) | GASIM · CoraFull, Arxiv, Reddit | mean AP | Hurt |
| Replace the factor-based router with a sequential router (ablation) | GASIM · CoraFull, Arxiv, Reddit | mean AP | Hurt |
| Replace the modular super-network with a vanilla super-network (ablation) | GASIM · CoraFull, Arxiv, Reddit | mean AP | Hurt |
Surprising findings
- DARTS achieves high average performance on Arxiv (61%) while forgetting heavily (AF -16.6%), indicating that searching a better architecture per task can harm past tasks.
- Existing GNAS methods do not consistently beat hand-designed GNNs in the continual setting.
Limitations
- Focuses only on homogeneous graphs; heterogeneous graphs left for future work.
- Relies on hyperparameters K (number of modules) and lambda (invariance trade-off).
Experimental setup
Datasets: CoraFull Arxiv Reddit
Benchmarks: CoraFull Arxiv Reddit
Metrics: average performance (AP) mean average performance average forgetting (AF)
Baselines: GCN GAT GIN GraphSage GraphConv SGC Random search GraphNAS GASSO DARTS
Builds on: DARTS GASSO continual graph learning (CGL) replay/condensation methods distribution matching for latent factors invariant risk minimization
Closest papers by meaning
- Unsupervised Graph Neural Architecture Search with Disentangled Self-Supervision (NeurIPS 2023 · similarity 0.91)
- Exploring Rationale Learning for Continual Graph Learning (AAAI 2025 · similarity 0.87)
- AutoGFM: Automated Graph Foundation Model with Adaptive Architecture Customization (ICML 2025 · similarity 0.87)
- Boosting Order-Preserving and Transferability for Neural Architecture Search: a Joint Architecture Refined Search and Fine-tuning Approach (CVPR 2024 · similarity 0.86)
- Efficient Graph Continual Learning via Lightweight Graph Neural Tangent Kernels-based Dataset Distillation (ICML 2025 · similarity 0.86)
- Neural Architecture Retrieval (ICLR 2024 · 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.