Threshold & Category Analysis

Category-specific alert thresholds for heterogeneous urban service demand

Urban service-demand forecasting is heterogeneous. Noise, housing, public safety, water/sewer, sanitation, infrastructure, environment, parking/traffic, and other categories have different base rates and alertability. Category-specific thresholds make this heterogeneity explicit.

Threshold Selection Summary

Score Column	Model	Threshold Mode	Selected Threshold	Val F1	Val Precision	Val Recall	Val ROC-AUC
lgbm_score	Tuned LightGBM	global	0.1670	0.3713	0.2853	0.5313	0.7547
xgb_score	Tuned XGBoost	global	0.1622	0.3683	0.2720	0.5699	0.7539
ensemble_lgbm_w0p500_score	Ensemble LGBM(0.50) + XGB(0.50)	global	0.1691	0.3720	0.2838	0.5396	0.7563

Category Threshold Comparison

Thresholds vary meaningfully across categories — reflecting different base rates and signal strengths

Category Performance (F1, Precision, Recall)

Alert Rate vs Base Rate by Category

Category Support Table

Category	Rows	Positives	Base Rate	Threshold (Ensemble)	F1	Precision	Recall	ROC-AUC
Noise	27,510	3,982	14.47%	0.1700	0.4423	0.3446	0.6173	0.7952
Housing & Building	27,510	3,720	13.52%	0.1901	0.4049	0.3476	0.4849	0.7803
Public Safety	27,510	3,536	12.85%	0.1571	0.3968	0.3157	0.5339	0.7812
Water/Sewer	27,510	3,462	12.58%	0.2050	0.3964	0.3319	0.4919	0.7822
Other	27,510	3,620	13.16%	0.1550	0.3725	0.2849	0.5381	0.7557
Parking/Traffic	27,510	3,586	13.04%	0.1650	0.3644	0.2810	0.5181	0.7470
Environment	27,510	2,896	10.53%	0.1576	0.3585	0.2665	0.5477	0.7830
Infrastructure	27,510	3,596	13.07%	0.1879	0.3495	0.2764	0.4753	0.7297
Sanitation	27,510	3,172	11.53%	0.1773	0.3194	0.2625	0.4076	0.7300