Research map / Graph learning / Graph self-supervised learning
Multimodal Graph Neural Architecture Search under Distribution 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 OMG-NAS, a reinforcement-learning multimodal graph NAS that decorrelates multimodal graph features via learned sample weights and shares those weights across searched architectures to improve out-of-distribution generalization.
Problem
Multimodal graph neural architecture search assumes identically distributed train and test data, so searched architectures overfit spurious correlations and generalize poorly under distribution shifts in multimodal graphs.
Approach
Combines multimodal graph feature decorrelation (random Fourier features with learned intra-modal sample weights, plus OGM-GE gradient balancing) with a global sample weight estimator whose weights are warm-started across successive sampled architectures, all inside an RNN-controller RL search loop.
Key results
- On Recipe Multi-OOD, OMG-NAS reaches 75.70±2.48% accuracy versus 68.84±4.52% for MG-NAS.
- On Tencent Multi-OOD, OMG-NAS achieves 66.82±1.35% versus 64.25±3.45% for MG-NAS.
- On Amazon review Multi-OOD-S, OMG-NAS achieves 88.40±2.13% versus 85.13±4.68% for MG-NAS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use OMG-NAS instead of MG-NAS | MG-NAS · Tencent Multi-OOD | accuracy: 64.25±3.45 -> 66.82±1.35 | Helped |
| use OMG-NAS instead of MG-NAS | MG-NAS · Amazon review Multi-OOD | accuracy: 85.13±4.68 -> 88.40±2.13 | Helped |
| use OMG-NAS instead of MG-NAS | MG-NAS · Amazon review Multi-OOD-Shoes, Baby, Dress, and Trousers | accuracy: 86.75±3.32 -> 88.47±2.82; 64.97±4.97 -> 68.46±1.27; 68.55±4.12 -> 71.65±1.93; 68.22±8.80 -> 75.56±5.41 | Helped |
| use OMG-NAS instead of MG-NAS | MG-NAS · Recipe Multi-OOD | accuracy: 68.84±4.52 -> 75.70±2.48 | Helped |
| use OMG-NAS instead of MG-NAS | MG-NAS · Recipe Single-OOD | accuracy: 75.82±4.12 -> 76.53±2.73 | Helped |
| add multimodal graph feature decorrelation (MGFD) rather than single feature decorrelation (SFD) (ablation) | fixed GCN, GAT, and MGAT · Tencent, Amazon review, and Recipe Multi-OOD | accuracy: GCN: 61.00±1.25 -> 55.26±2.24; 52.80±9.20 -> 61.85±3.51; 61.65±5.05 -> 70.63±2.10; MGAT: 60.06±7.99 -> 65.66±1.41; 56.55 | Mixed |
| add MGFD to a fixed GCN (ablation) | GCN · Tencent Multi-OOD | accuracy: 55.88±0.96 -> 55.26±2.24 | No clear effect |
| add MGFD to a fixed MGAT (ablation) | MGAT · Amazon review Multi-OOD | accuracy: 53.60±7.01 -> 58.30±4.97 | Helped |
| transfer global sample weights across architectures (ablation) | OMG-NAS · Tencent, Amazon review, and Recipe Multi-OOD | accuracy: 62.54±3.11 -> 66.82±1.35; 63.28±6.53 -> 71.65±1.93; 72.85±3.54 -> 75.70±2.48 | Helped |
| transfer sample weights from GCN to GAT (ablation) | Recipe architecture search · Recipe | accuracy: +7.22% | Helped |
| transfer sample weights from GAT to GCN (ablation) | Recipe architecture search · Recipe | accuracy: -5.80% | Hurt |
Surprising findings
- Applying OOD-GNN sample reweighting to fixed MGNNs often does not improve and sometimes lowers accuracy, e.g., GAT with OOD-GNN drops on Amazon Multi-OOD-T.
- Transferring sample weights across architectures does not always help; GAT to GCN transfer drops accuracy by 5.80% in Table 5.
Limitations
- Evaluated on only three multimodal graph datasets with text and image modalities.
- Relies on OGM-GE and RFF-based decorrelation, adding design complexity to the search loop.
Experimental setup
Datasets: Tencent WeChat article graph Amazon review multimodal graph Recipe multimodal graph
Benchmarks: Tencent Amazon review Recipe
Metrics: classification accuracy
Baselines: GCN GAT MGAT Mixup DIR CIGA EERM OOD-GNN Random Search GraphNAS MG-NAS
Builds on: MG-NAS GraphNAS OGM-GE OOD-GNN random Fourier features
Closest papers by meaning
- NAS-Bench-Graph: Benchmarking Graph Neural Architecture Search (NeurIPS 2022 · similarity 0.87)
- Graph Neural Architecture Search Under Distribution Shifts (ICML 2022 · similarity 0.87)
- One-shot Graph Neural Architecture Search with Dynamic Search Space (AAAI 2021 · similarity 0.87)
- Mosaic of Modalities: A Comprehensive Benchmark for Multimodal Graph Learning (CVPR 2025 · similarity 0.87)
- AutoGFM: Automated Graph Foundation Model with Adaptive Architecture Customization (ICML 2025 · similarity 0.85)
- Data-Augmented Curriculum Graph Neural Architecture Search under Distribution Shifts (AAAI 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.