Research map / Graph learning / GNN architectures
Identifying Guarantors of War Veterans Using Robust-SEAL: A Case of the Korean War
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 Robust-SEAL, combining robust GCN (variance-based down-weighting) with SEAL's labeling trick, to predict missing combined-operation links among Korean War army units to recommend guarantors for veterans' buddy statements.
Problem
Veterans lacking military records struggle to find comrades to vouch for their service (buddy statements); combined-operation war records are incomplete and noisy, so guarantor candidates are limited.
Approach
Combine a probabilistic Gaussian robust GCN (variance-based attention that down-weights high-uncertainty node features) with SEAL's subgraph and double-radius node labeling to capture global context, and apply the hybrid to link prediction on a Korean War unit network.
Key results
- Transductive nDCG@3: Robust-SEAL 0.7238 vs SEAL 0.7031 and Robust GCN 0.6617.
- Semi-inductive nDCG@3: Robust-SEAL 0.7342 vs SEAL 0.6945 and GCN 0.5810.
- Under Gaussian noise (N(0,100^2)) and 20% edge deletion, Robust-SEAL nDCG@3 0.713 and 0.692, above SEAL (0.649, 0.607).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine robust GCN with SEAL | SEAL link prediction · Korean War, transductive | nDCG@3: 0.7031(0.0209) -> 0.7238(0.0182) | Helped |
| Combine robust GCN with SEAL | SEAL link prediction · Korean War, semi-inductive | nDCG@3: 0.6945(0.0208) -> 0.7342(0.0251) | Helped |
| Combine robust GCN with SEAL | SEAL link prediction · Korean War, semi-inductive, casualty noise N(0, 10²) | nDCG@3: 0.662(0.021) -> 0.718(0.022) | Helped |
| Combine robust GCN with SEAL | SEAL link prediction · Korean War, semi-inductive, casualty noise N(0, 100²) | nDCG@3: 0.649(0.023) -> 0.713(0.021) | Helped |
| Combine robust GCN with SEAL | SEAL link prediction · Korean War, semi-inductive, 5% edge deletion | nDCG@3: 0.673(0.022) -> 0.723(0.016) | Helped |
| Combine robust GCN with SEAL | SEAL link prediction · Korean War, semi-inductive, 20% edge deletion | nDCG@3: 0.607(0.028) -> 0.692(0.017) | Helped |
Surprising findings
- Robust GCN gains were larger under perturbations (11-25%) than on clean data (about 9%).
Limitations
- Casualty data available only at regiment level, so unit-level node features are noisy.
- Evaluated on a single war dataset; other wars could not be tested due to data availability.
- Fairness concern: prediction accuracy biased toward high-degree nodes.
Experimental setup
Datasets: Korean War combined operation records (Cho et al. 2017; Min 2009) War Memorial of Korea casualty records
Benchmarks: Korean War combined operation link prediction (transductive and semi-inductive)
Metrics: nDCG@K (K=1-5)
Baselines: GCN GraphSAGE GAT GraphSAINT Robust GCN SEAL LGLP
Builds on: SEAL Robust GCN (RGCN) GCN LGLP
Closest papers by meaning
- CEGA: A Cost-Effective Approach for Graph-Based Model Extraction and Acquisition (ICML 2025 · similarity 0.79)
- Bounding the Expected Robustness of Graph Neural Networks Subject to Node Feature Attacks (ICLR 2024 · similarity 0.79)
- ROK-FORTRESS: Measuring the Effect of Geopolitical Transcreation for National Security and Public Safety (NeurIPS 2026 · similarity 0.78)
- Mixture of Weak and Strong Experts on Graphs (ICLR 2024 · similarity 0.77)
- Adversarial Robust Generalization of Graph Neural Networks (ICML 2025 · similarity 0.77)
- On provable privacy vulnerabilities of graph representations (NeurIPS 2024 · similarity 0.77)
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.