GOURMET v31.5: GNN Enhancement Report

Generated: 2026-06-11T10:50:57.939237+00:00 Device: cuda

Results

What Changed

1. Entity features: Added 384-dim KG embeddings (from Stream A) as node features
2. Temporal labels: Used temporal edges as additional supervision signal
3. Cross-domain prediction: Entity→domain link prediction instead of just observed_in
4. Architecture: MLP with 256→128→64 hidden layers, dropout regularization

Data

• Entity embeddings: 9665 (384-dim from all-MiniLM-L6-v2)
• Domain embeddings: 11 (384-dim)
• Training examples: 19330 (balanced positive/negative)

v31.5 Stream E — GNN Enhancement. 2026-06-11T10:50:57.939267+00:00