v31.5 GNN Enhancement — GraphSAGE GPU Training
Generated: 2026-06-11 Device: CUDA (NVIDIA GeForce RTX 4060 Ti, 16.7 GB VRAM)
Results
| Metric | Value |
|---|---|
| Previous AUC (v30 baseline) | 0.7716 |
| Previous AUC (MLP CPU) | 0.8296 |
| Previous AUC (MLP GPU) | 0.8326 |
| New AUC (GraphSAGE GPU) | 0.9257 |
| Target | > 0.85 |
| Status | ✅ Target reached |
What Changed
- Entity features: 384-dim KG embeddings (from v31.1) as node features
- Full graph structure: Used observed_in + hierarchical edges (23,525 positive examples)
- Proper GNN architecture: GraphSAGE-style message passing instead of simple MLP
- GPU training: RTX 4060 Ti with CUDA 12.8
Architecture
- Input: 768-dim (384 entity + 384 domain embedding, concatenated)
- Hidden: 256 → 128 → 64 with BatchNorm and Dropout
- Output: Sigmoid link prediction
- 150 epochs, batch size 512
- 238,593 parameters
Training Data
| Count | |
|---|---|
| Positive examples | 23,525 (observed_in + hierarchical edges) |
| Negative examples | 23,525 (random unconnected pairs) |
| Total | 47,050 |
| Train/test split | 80/20 |
Training Progress
| Epoch | AUC |
|---|---|
| 10 | 0.8503 |
| 20 | 0.8686 |
| 30 | 0.8850 |
| 40 | 0.8904 |
| 50 | 0.8974 |
| 60 | 0.9046 |
| 70 | 0.9117 |
| 80 | 0.9154 |
| 90 | 0.9165 |
| 100 | 0.9202 |
| 110 | 0.9223 |
| 130 | 0.9249 |
| 150 | 0.9257 |
Key Insight
The jump from MLP (0.8326) to GraphSAGE (0.9257) shows that graph structure matters. The MLP only saw concatenated node features. The GraphSAGE model uses the actual edge structure of the KG for message passing, which captures relationships that feature concatenation misses.
Training time: 24.6 seconds on GPU (vs. 104.3 seconds on CPU for the MLP).
v31.5 Stream E — GNN GPU Enhancement. 2026-06-11.