Padjective Tag Hierarchy

Machine learning insights into Shopify product tag organization

Data sourced from cantbuymelove.industrial-linguistics.com powering Shopify taxonomy classification and filtered to taxonomies with at least five products.

Last updated 2026-11-17 01:40 UTC

25,838 Products used
1,084 Taxonomies covered
22,587 Tags used
55,537 Total tags
9,556 Tag battles

Dataset coverage

Training data spans 25,838 products across 1,084 taxonomies. Of 55,537 total tags in the dataset, 22,587 tags were used (tags appearing fewer than 5 times were filtered out). 11,570 products were discarded due to missing or sparse taxonomy labels. Explore the full dataset → | View defective taxonomy labels →

Dummy Baseline

Always predicts most common taxonomy (baseline for comparison)

0.6258 Avg p-adic loss
1 Parameter
View model →

Importance-Optimised p-adic Linear Regression

P-adic coefficients assigned to tags to predict taxonomy

0.3718 Avg p-adic loss
2,498 Avg non-zero coefficients
View model →

Level-wise Logistic Regression

Hierarchy-aware top-down classifier that always emits a valid taxonomy path

0.1008 Avg p-adic loss
83.01% Prefix-2 accuracy
132,415 Non-zero params
View model →

Zubarev Regression (UMLLR init)

Stochastic p-adic optimization starting from UMLLR (arXiv:2503.23488)

0.4257 Avg p-adic loss
5,009 Non-zero coefficients
View fold details →

Zubarev Regression (Zeros init)

Stochastic p-adic optimization starting from zeros (arXiv:2503.23488)

0.4168 Avg p-adic loss
4,061 Non-zero coefficients
View fold details →

Zubarev Mahler-1 (UMLLR init)

Mahler affine basis (degree 1) with UMLLR initialization

0.4251 Avg p-adic loss
4,694 Non-zero coefficients
View fold details →

Zubarev Mahler-2 (UMLLR init)

Mahler quadratic basis (degree 2) with UMLLR initialization

0.4245 Avg p-adic loss
4,698 Non-zero coefficients
View fold details →

Unconstrained Logistic Regression

L1-regularized model using ALL tags

0.2878 Avg p-adic loss
11,886 Non-zero params
View model →

Decision Tree

Unconstrained tree using ALL tags

0.2219 Avg p-adic loss
137,177 Effective params
View model →

Unconstrained Neural Network

L1-regularized NN with weight pruning

0.2574 Avg p-adic loss
22,260 Non-zero params
View model →

Parameter Constrained Neural Network

Neural network predicting taxonomy from tags

0.6058 Avg p-adic loss
864 Avg input weights
View model →

Parameter Constrained Logistic Regression

Logistic regression model predicting Shopify taxonomy from tags

0.8304 Avg p-adic loss
34,406 Avg parameters
View model →

ELO-Inspired Rankings

Battle-tested tag hierarchy from product title positions

9,556 Tag battles
View rankings →

Benchmark Comparisons

Dedicated `latest` and `paper` benchmark pages, including the average active parameters touched per classification for the importance-optimised p-adic linear regressor.

1.13 Latest active params / classification
1.11 Paper active params / classification
Open benchmark pages →

Taxonomy distribution

Taxonomy class distribution
Distribution of products across the most common taxonomy classes

Top 10 taxonomy classes

Taxonomy IDNamePathSamplesShare
gid://shopify/TaxonomyCategory/btBaby & Toddler43981.5%
gid://shopify/TaxonomyCategory/lbLuggage & Bags151090.4%
gid://shopify/TaxonomyCategory/buBundles51040.4%
gid://shopify/TaxonomyCategory/hgHome & Garden14900.3%
gid://shopify/TaxonomyCategory/gcGift Cards11510.2%
gid://shopify/TaxonomyCategory/paProduct Add-Ons19440.2%
gid://shopify/TaxonomyCategory/osOffice Supplies18420.2%
gid://shopify/TaxonomyCategory/sgSporting Goods23290.1%
gid://shopify/TaxonomyCategory/aaApparel & Accessories1250.1%
gid://shopify/TaxonomyCategory/aeArts & Entertainment3250.1%

Tags with strongest signal

TagTop taxonomyWeightMax |weight|
No tag signal data available

Historical Performance Trends

Tracking model performance and dataset growth over time. Lower p-adic loss indicates better predictions.

Historical model performance trends
Model performance vs number of products
Model Slope (per product) Intercept R² p-value
Importance-Optimised p-adic LR0.0000010.34050.05934.05e-06
PCLR0.0000100.61520.40642.60e-41
PCNN0.0000000.60910.00000.9831
ULR0.0000030.21910.57335.92e-61
UNN0.0000010.21900.07556.18e-07
Decision Tree0.0000020.18960.33982.47e-30
Zubarev (UMLLR)0.0000020.39260.35901.71e-31
Zubarev (zeros)-0.0000010.45780.02120.0104
Zubarev (M1)0.0000010.39730.26194.99e-22
Zubarev (M2)0.0000010.39550.29504.10e-25
Dummy Baseline-0.0000100.81870.31089.37e-29

Extrapolation Analysis: When Will Importance-Optimised p-adic LR Outperform Other Models?

Based on current regression trends, we can extrapolate when Importance-Optimised p-adic LR will achieve better performance (lower p-adic loss) than other models as the dataset grows. The confidence intervals are calculated using bootstrap resampling (n=1000).

Model Crossover Point
(products)
95% Confidence Interval Probability Estimated Date
UNN (Unconstrained Neural Networks)314,146139,510 - 2,639,566 (95% CI, σ=2,479,708)>95%2042-08-26 (±uncertain, R²=0.615, growth=51.0/product/day)
ULR (Unconstrained Logistic Regression)62,68554,891 - 74,577 (95% CI, σ=5,010)>95%2029-02-20 (±uncertain, R²=0.615, growth=51.0/product/day)
Decision Tree196,346133,804 - 368,405 (95% CI, σ=71,902)>95%2036-04-27 (±uncertain, R²=0.615, growth=51.0/product/day)

Statistical Notes: The crossover points are calculated by finding where the regression lines intersect. The 95% confidence intervals are derived from bootstrap resampling of the regression parameters. The probability estimates indicate the likelihood that the crossover will occur given the current trends. Date predictions are based on linear extrapolation of dataset growth and should be interpreted with caution.

Model performance vs number of distinct tags
Model Slope (per tag) Intercept R² p-value
Importance-Optimised p-adic LR0.0000010.33530.04863.19e-05
PCLR0.0000170.49020.59374.83e-70
PCNN0.0000070.51330.13974.80e-13
ULR0.0000050.18140.83392.12e-126
UNN0.0000050.16190.59249.75e-64
Decision Tree0.0000030.15920.71691.37e-88
Zubarev (UMLLR)0.0000000.40700.00540.1957
Zubarev (zeros)-0.0000020.47730.06733.83e-06
Zubarev (M1)-0.0000000.41690.00800.1171
Zubarev (M2)-0.0000000.41370.00160.4903
Dummy Baseline-0.0000030.74780.01550.0228

Extrapolation Analysis: When Will Importance-Optimised p-adic LR Outperform Other Models?

Based on current regression trends, we can extrapolate when Importance-Optimised p-adic LR will achieve better performance (lower p-adic loss) than other models as the dataset grows. The confidence intervals are calculated using bootstrap resampling (n=1000).

Model Crossover Point
(tags)
95% Confidence Interval Probability Estimated Date
UNN (Unconstrained Neural Networks)44,75939,613 - 51,286 (95% CI, σ=3,010)>95%2028-02-27 (±uncertain, R²=0.998, growth=47.0/tag/day)
ULR (Unconstrained Logistic Regression)40,62837,346 - 44,696 (95% CI, σ=1,886)>95%2027-12-02 (±uncertain, R²=0.998, growth=47.0/tag/day)
Decision Tree75,40963,743 - 94,364 (95% CI, σ=7,748)>95%2029-12-10 (±uncertain, R²=0.998, growth=47.0/tag/day)

Statistical Notes: The crossover points are calculated by finding where the regression lines intersect. The 95% confidence intervals are derived from bootstrap resampling of the regression parameters. The probability estimates indicate the likelihood that the crossover will occur given the current trends. Date predictions are based on linear extrapolation of dataset growth and should be interpreted with caution.

Model complexity vs performance (parameter count vs p-adic loss)
Both axes use log scale. The red line is the fixed parsimoniousness baseline rather than a fitted regression.

Why parsimony matters. The question here is not just which model has the lowest loss, but which model gets good p-adic loss with the fewest effective parameters. That is exactly where the smaller p-adic models are interesting.

Where this baseline came from. The original score came from a log-log regression on model size versus loss, rounded to -0.1 × log₁₀(params) - 0.2. Looking across historical snapshots, those scores drifted as the dataset covered more taxonomies, so the current baseline adds + 0.3 × log₁₀(taxonomies / 1,000) to keep comparisons stable as the benchmark grows. For readability, we also re-centre the displayed score by dropping the old constant offset; that keeps the current tables mostly positive without changing the relative comparisons.

Parsimoniousness baseline: log₁₀(loss) = -0.1 × log₁₀(params) + 0.3 × log₁₀(taxonomies / 1,000)
Current snapshot taxonomies: 1,084
Parsimony score = baseline log₁₀(loss) − observed log₁₀(loss). Positive means better than baseline.

Model Params Loss log₁₀(params) log₁₀(loss) Baseline log₁₀(loss) Parsimony score
Level-wise Logistic132,4150.10085.1219-0.9966-0.5017+0.4950
Dummy10.62580.0000-0.20360.0105+0.2141
UNN22,2600.25744.3475-0.5895-0.4242+0.1652
Decision Tree137,1770.22195.1373-0.6539-0.5032+0.1507
ULR11,8860.28784.0750-0.5409-0.3970+0.1439
Importance-Optimised2,4980.37183.3977-0.4297-0.3293+0.1004
Zubarev (zeros)4,0610.41683.6087-0.3801-0.3504+0.0297
Zubarev (M2)4,6980.42453.6719-0.3721-0.3567+0.0155
Zubarev (M1)4,6940.42513.6715-0.3716-0.3566+0.0149
Zubarev (UMLLR)5,0090.42573.6997-0.3709-0.3595+0.0114
PCNN8640.60582.9365-0.2176-0.2831-0.0655
PCLR34,4060.83044.5366-0.0807-0.4432-0.3624
Historical parsimony score stability
Left: parsimony score versus dataset size. Right: score distribution across historical snapshots. Positive means better than the taxonomy-adjusted baseline.
Model Snapshots Mean score Std dev Span Latest score Latest products
Dummy Baseline335+0.07530.12300.3911+0.214125,838
Unconstrained Neural Network with L1319+0.11390.05980.2763+0.165225,838
Decision Tree288+0.13860.02410.1372+0.150725,838
Unconstrained Logistic Regression with L1321+0.15390.02430.1836+0.143925,838
Importance-Optimised $p$-adic Linear Regression288+0.07700.03950.1295+0.100425,838
Zubarev (UMLLR init)309-0.03740.03170.1192+0.011425,838
PCNN319-0.17020.08380.2645-0.065525,838
PCLR319-0.38020.02560.2073-0.362425,838

Smaller standard deviation and span mean a model’s parsimoniousness is more stable as the dataset grows.

Unconstrained models: complexity vs performance (log-log scale)
Unconstrained models only (no PCLR/PCNN). Both axes on log scale.

Regression: log₁₀(loss) = slope × log₁₀(params) + intercept

Slope Intercept R² p-value Significant? n
-0.0871 -0.1882 0.9675 0.0025 Yes 5
Model performance trajectory over time
Arrows show how each model's complexity and performance have changed over time.