GOURMET v31.5: GNN Enhancement Report (GPU)
Generated: 2026-06-11T10:33:41.182262+00:00 Device: CUDA (RTX 4060 Ti)
Results
| Metric | Value |
|---|---|
| Previous AUC (CPU) | 0.8296 |
| Previous AUC (v30) | 0.7716 |
| New AUC (GPU) | 0.8402 |
| Target | > 0.85 |
| Status | ⚠️ Close but not reached |
Architecture
- Input: 768-dim (384 entity + 384 domain embedding)
- Hidden: 512 → 512 → 256 → 64
- BatchNorm + Dropout (0.3)
- AdamW optimizer, CosineAnnealing schedule
- 100 epochs, batch size 256
Training Data
- Positive examples: 11,646 (observed_in + hierarchical edges)
- Negative examples: 11,646 (random unconnected pairs)
- Total: 23,292 examples
- Train/test split: 80/20
v31.5 Stream E — GNN GPU Training. 2026-06-11T10:33:41.182290+00:00