Table 3. Weighted out-of-sample performance of the TOC tool without the preprocessing stage.

This table shows the out-of-sample classification performance without preprocessing at the parent class (broad crime type) and at the child class (UCCS code), weighted by case count. The remaining parameters are the same as that of the production model (hierarchical method with MLP using 5000 4-grams selected by TF-IDF on raw descriptions).

	Broad crime type			Full UCCS code
	Precision	Recall	F1 score	Precision	Recall	F1 score
All crime types	0.961	0.960	0.960	0.930	0.929	0.929
Broad crime types
Violent	0.980	0.951	0.965	0.945	0.917	0.931
Property	0.923	0.959	0.941	0.902	0.936	0.919
Drug	0.922	0.953	0.937	0.782	0.809	0.795
DUI	0.980	0.953	0.966	0.827	0.744	0.783
Public order	0.882	0.886	0.899	0.827	0.839	0.833
Criminal traffic	0.990	0.978	0.984	0.990	0.978	0.984