GOURMET v32.5: GNN Enhancement Report (GPU Training)
Generated: 2026-06-12T07:45:30.844453+00:00 Device: CUDA (RTX 4060 Ti) Training time: 30.7s
Results
| Metric | Value |
|---|---|
| v31 Baseline AUC | 0.9257 |
| v32.5 AUC | 0.9987 |
| Improvement | +0.0730 |
| Status | ✅ IMPROVED |
Architecture
- Entity encoder: Linear(384→512) + BN + ReLU + Dropout + ResidualBlock(512)
- Domain encoder: Linear(384→512) + BN + ReLU + Dropout + ResidualBlock(512)
- Interaction: element-wise product of entity/domain hidden states
- Predictor: Linear(1536→512) + ResidualBlock → 256 → 64 → 1
- Total params: 2,930,689
- Dropout: 0.3
Training
- Epochs: up to 200 (early stopping patience=20)
- Batch size: 512
- Optimizer: AdamW (lr=0.001, weight_decay=0.0001)
- Scheduler: CosineAnnealingWarmRestarts
- Loss: BCEWithLogitsLoss
- Train/val split: 80/20
Data
- Positive examples: observed_in + hierarchical edges (entity-domain pairs with embeddings)
- Negative examples: 100% random + 50% hard negatives
- Feature dim: 1536 (entity_emb + domain_emb + product + abs_diff)
Training Log (last 10 entries)
| Epoch | Train Loss | Val Loss | Val AUC | Val AP |
|---|---|---|---|---|
| 1 | 0.4111 | 1.5905 | 0.7422 | 0.8535 |
| 5 | 0.0267 | 0.0548 | 0.9980 | 0.9988 |
| 10 | 0.0075 | 0.0571 | 0.9985 | 0.9992 |
| 15 | 0.0027 | 0.0598 | 0.9987 | 0.9993 |
| 20 | 0.0023 | 0.0625 | 0.9986 | 0.9992 |
| 25 | 0.0100 | 0.0786 | 0.9975 | 0.9986 |
| 30 | 0.0069 | 0.0699 | 0.9980 | 0.9986 |
| 35 | 0.0045 | 0.0803 | 0.9980 | 0.9987 |
v32.5 Stream E — GNN GPU Enhancement. 2026-06-12T07:45:30.844695+00:00