taxonomy_association
Taxonomy-peaked tags first
For each tag, measure the share of its training occurrences that land in its single most common taxonomy. Tags with the strongest one-taxonomy association are scored first.
← Back to Paper benchmark summary
Fixed benchmark snapshot shared by the Hugging Face notebook, benchmark-only HTML build, and generated TeX includes. Best strategy in this bundle: taxonomy_association (0.263237 mean p-adic loss).
For these UMLLR ablations, Avg active params / classification is the mean number of active coefficients touched while classifying one product.
Ordering methods
The ablation keeps the greedy p-adic regressor fixed and changes only the tag ordering heuristic used before coefficient fitting.
random
Seeded random control
Uses a seeded random shuffle of the training tag vocabulary as a control condition.
frequency
Most common tags first
Ranks tags by how often they appear in the training products.
battle_elo
Pairwise battle ranking
Ranks tags by fold-local Elo scores estimated from tag battles, while excluding the holdout fold from the ranking fit.
mean_title_position
Average title position
Ranks tags by their average recorded title position in the training products.
| Strategy | Mean p-adic loss | Δ vs battle_elo | Fold wins | Exact acc. | Prefix-2 acc. | Avg active params / classification |
|---|---|---|---|---|---|---|
| taxonomy_association | 0.263237 | -0.061688 | 5/5 | 49.82% | 65.38% | 1.11 |
| random | 0.312856 | -0.012069 | 21/25 | 43.73% | 59.89% | 1.41 |
| frequency | 0.319671 | -0.005254 | 4/5 | 42.88% | 60.20% | 1.91 |
| battle_elo | 0.324925 | 0.000000 | 0/5 | 43.63% | 59.75% | 1.42 |
| mean_title_position | 0.342006 | 0.017081 | 1/5 | 42.28% | 58.67% | 1.68 |