Training data spans 6,108 products across
368 taxonomies.
Of 24,460 total tags in the dataset,
9,365 tags were used
(tags appearing fewer than 5 times were filtered out).
4,184 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)
Arts & Entertainment > Hobbies & Creative Arts > Collectibles
3.2.2
79
1.3%
Tags with strongest signal
Tag
Top taxonomy
Weight
Max |weight|
FRAMED ARTWORK
3.2.2
5.8150
5.8150
BLUE
14.11.10.4.3
5.4904
5.4904
WOMENS
1.8.7
5.4341
5.4341
ACCESSORIES
1.2.4
5.2483
5.2483
GIFT
14.15.1.9
5.0948
5.0948
WHOLESALE
14.11.10.7.9
5.0544
5.0544
VEGAN
13.3.5.2
5.0324
5.0324
KIDS
13.1.20
4.9532
4.9532
NEW ARRIVALS
13.3.2.8.4
4.8409
4.8409
PLUS SIZE
1.1.1.1.5
4.8031
4.8031
Historical Performance Trends
Tracking model performance and dataset growth over time. Lower p-adic loss indicates better predictions.
Model
Slope (per product)
Intercept
R²
p-value
Importance-Optimised p-adic LR
0.000012
0.2991
0.3005
8.13e-08
PCLR
0.000082
0.2835
0.7514
3.32e-26
PCNN
0.000082
0.2458
0.8471
8.97e-35
ULR
0.000009
0.1766
0.2355
2.00e-04
UNN
0.000027
0.0671
0.7473
1.50e-16
Decision Tree
0.000009
0.1401
0.2973
3.52e-05
Zubarev (UMLLR)
0.000020
0.3037
0.8125
4.01e-16
Zubarev (zeros)
0.000027
0.2936
0.8331
3.85e-17
Zubarev (M1)
0.000007
0.3744
0.3632
2.41e-05
Zubarev (M2)
0.000011
0.3547
0.5395
3.08e-08
Dummy Baseline
-0.000073
1.0940
0.5802
4.63e-14
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).
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
Slope (per tag)
Intercept
R²
p-value
Importance-Optimised p-adic LR
0.000014
0.2424
0.3331
1.12e-08
PCLR
0.000090
-0.0697
0.7355
4.17e-25
PCNN
0.000090
-0.1065
0.8234
3.11e-32
ULR
0.000009
0.1442
0.2658
6.61e-05
UNN
0.000027
-0.0256
0.7661
2.15e-17
Decision Tree
0.000009
0.1069
0.3271
1.16e-05
Zubarev (UMLLR)
0.000021
0.2276
0.8448
8.96e-18
Zubarev (zeros)
0.000028
0.1928
0.8509
4.01e-18
Zubarev (M1)
0.000008
0.3465
0.3708
1.87e-05
Zubarev (M2)
0.000011
0.3161
0.5375
3.36e-08
Dummy Baseline
-0.000078
1.3849
0.6068
5.22e-15
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).
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.
Parameter count (log scale) vs p-adic loss. Sparse models use fewer non-zero parameters.
Regression: p-adic loss = slope × log₁₀(params) + intercept
Line
Slope
Intercept
R²
p-value
Significant?
n
With Dummy
-0.0698
0.6474
0.2301
0.1354
No
11
Without Dummy
-0.1223
0.8396
0.1616
0.2495
No
10
Unconstrained models only (no PCLR/PCNN). Both axes on log scale.
