MeshGraph GAT
MeshGraph GAT applies four-head graph attention to the same 600-node enterprise compromise benchmark used by MeshGraph GCN. Validation and test labels remain held out in the transductive graph.
The report compares GAT with the previously trained GCN and feature-only logistic control, then audits normalized attention entropy and the highest-weight test edges. The Space exposes neighbor attention for every node.
Verified local result
The 772-parameter GAT reached 0.8412 ROC-AUC, 0.6316 average precision, and 0.5965 F1. The 802-parameter GCN remained stronger at 0.8564/0.6976/0.6333. Mean normalized GAT attention entropy was 0.929, indicating broadly distributed rather than sharply selective neighbor weighting.
uv run python projects/meshgraph-gat/train.py
uv run pytest tests/test_meshgraph_gat.py
Hosted showcase
This free static Space preserves the complete original Gradio source, trained artifacts, evaluation files, and local launch requirements. Hugging Face now requires PRO for CPU-backed Gradio hosting, so the public landing page is static while the checked-in app.py remains the authoritative runnable demo.