PCNN Fold 4

Back to PCNN overview · Back to main index

Fold metrics

Metric	Value
Test accuracy	14.10%
Test F1 score	0.1101
Hierarchical loss	0.78396950
P-adic loss (total)	1258.43860000
P-adic loss (mean)	0.68505090
Prime base	79
Hidden layer size	27
Max tags	32
Training samples	7,273
Test samples	1,837

P-adic loss breakdown

Agreement	Count	Share	Cost per mistake	Total contribution
Exact match	259	14.10%	0.000000	0.000000
p^4	2	0.11%	0.000000	0.000000
p^3	78	4.25%	0.000002	0.000158
p^2	129	7.02%	0.000160	0.020670
p^1	112	6.10%	0.012658	1.417722
p^0	1,257	68.43%	1.000000	1257.000000

Tag Rank vs First-Layer Weight Magnitude

Tag rank vs max first-layer weight magnitude — Scatter plot showing the relationship between tag battle ranking and maximum absolute first-layer weight value across all hidden units. Shows which input features contribute most to the parameter constrained neural network's hidden representations.

About p-adic loss

P-adic loss measures the distance between predicted and true taxonomy using p-adic metric (base 79). Lower values indicate closer predictions in the taxonomy hierarchy. This metric is shared with the umllr model for comparison.