Table 1. Weighted out-of-sample performance of the TOC tool by broad crime type.

This table shows the out-of-sample classification performance of the production Text-based Offense Classification (TOC) model at the parent class (broad crime type) and at the child class [Uniform Crime Classification Standard (UCCS) code] weighted by the case count of each offense description. The model uses hierarchical classification method with multilayer perceptron (MLP) classifier trained at each parent node using 5000 4-grams selected by term frequency–inverse document frequency (TF-IDF) from preprocessed descriptions.

	Broad crime type			Full UCCS code
	Precision	Recall	F1 score	Precision	Recall	F1 score
All crime types	0.983	0.983	0.983	0.963	0.963	0.963
Broad crime types
Violent	0.997	0.994	0.995	0.993	0.989	0.991
Property	0.927	0.990	0.957	0.884	0.944	0.913
Drug	0.999	0.960	0.979	0.862	0.828	0.845
DUI	0.987	0.986	0.986	0.942	0.941	0.941
Public order	0.993	0.938	0.965	0.977	0.923	0.949
Criminal traffic	0.987	0.991	0.989	0.986	0.991	0.988