HGCNN benchmark proof demo
This page shows the current simulated benchmark harness: 500 train rows, 1000 holdout/test rows, and 500 holdout queries with ideal answers. It compares AVRAI's HGCNN paths with five RecBole baselines using the same frozen fixture, then shows why the result is still shadow evidence instead of empirical promotion.
The trained HGCNN candidate scorer put the ideal simulated answer first on the 500 holdout queries.
The governed Actualizer readback kept the ideal answer in the top three while applying safety and decision gating.
Best RecBole baseline on NDCG@3 for this exported simulated fixture.
Strong graph collaborative filtering baseline, but lower top-three coverage than the HGCNN paths.
This model did not fit the fixture shape well in the external run.
Classic pairwise baseline included as a low-complexity control.
Translation-style KG recommendation underperformed on this simulated ranking task.
NDCG@3 0.6415
trained shadow NDCG@3
governed readback NDCG@3
Full harness path from simulated packets to blocked decision receipt.
Creates unique packet-shaped rows with ideal answers, then checks temporal holdout and no raw payload storage.
Locks the 500 train, 1000 holdout, and 500 query split so every model sees the same task.
Translates the fixture into RecBole run material and verifies the selected baseline reports.
Runs the AVRAI HGCNN candidate scorer and Actualizer-bound readback without live authority.
Compares HGCNN and Actualizer metrics against BPR, LightGCN, KGNN-LS, KTUP, and KGIN.
Carries the result as shadow evidence and blocks empirical promotion until real validator-clean packet rows exist.