Back The Archive — the dataset behind these numbers

The Lab — offline evaluation

How the 0 recommenders compare on the held-out test set (per-user 80/20 split, k = 10). The point isn't one winner — it's the accuracy vs. beyond-accuracy trade-off. Best value per column is highlighted; hover a heading for its definition.

Highlighted row = the model powering the app (LTR hybrid + re-ranker) Click any row to drive a live “Top picks” rail
ModelP@10Recall@10NDCGCoverageDiversityNoveltySerendipity

User-user CF and the LTR hybrid lead on accuracy; the re-ranker, content-based and random lead on diversity / novelty. The hybrid's value is combining many candidate generators into one ranking, then the re-ranker trades a little precision for the diversity & serendipity the product cares about.