kg-link-prediction

Run the train_reference.py script with your NeuroOracle knowledge_graph.json, selecting a model via --model (complex, rgcn, graphsage, or gat). Set embedding dimension, epochs, and output directory; the script produces a checkpoint.pt and metrics.json.

R-GCN uses relation-specific graph convolution, GraphSAGE uses neighborhood aggregation, and GAT uses graph attention; all three pair with relation-aware DistMult decoding. Keep the same split seed when comparing encoders for fair evaluation.

No, the implementation uses native PyTorch operations and does not require PyTorch Geometric. You only need numpy, torch, and scikit-learn installed.

GNN models report AUROC, AUPRC, filtered MRR, and filtered Hits metrics. Validation and test edges are excluded from encoder adjacency, and negative samples are checked against all known positive triples.

Yes, pass --min-confidence with a threshold value to filter edges below that confidence level from the NeuroOracle graph before training. This helps control noise from weakly supported literature evidence.

Retrain whenever the graph snapshot changes materially. Record the graph hash and freeze year when embeddings are used for hindcasting, and verify no post-freeze edge enters training or message passing.