CRIMESENSEPOWERED BY CRIMENET
PRODUCTION ARCHITECTUREMODEL SYSTEM / NATIONAL INFERENCE
CrimeSense Risk Forecasting System
National spatial-temporal crime-risk forecasting powered by CrimeNet, a two-stage marked point-process system implemented with XGBoost intensity and conditional mark models.
NATIONAL SURFACETWO-STAGE XGBOOSTLIVE → +24H
01 / SYSTEM DEFINITION
CrimeNet is the infrastructure behind CrimeSense.
CrimeNet transforms historical public-safety records and national spatial, environmental, socioeconomic, temporal, and built-environment features into hourly H3 risk surfaces and conditional crime-category forecasts.
It spans data preparation, point-in-time feature generation, training construction, national inference, forecast materialization, adaptive map LOD, and production API serving. CrimeSense is the interactive application that makes those model states explorable.
02 / CRIMENET ARCHITECTURE
From event corpus to interactive forecast.
The historical corpus is expanded into a larger spatial-temporal training table with integration support; training examples and observed crimes are intentionally different quantities.
38 FEATURESλ(x,t)
How much modeled event activity is expected here?
An XGBoost Poisson / point-process intensity model estimates expected crime-event intensity for each canonical H3 cell under its spatial, temporal, environmental, and socioeconomic context.
- OUTPUT
- expected events / cell / hour
- OBJECTIVE
- point-process Poisson
- RUN ID
ff94186b570641e68eb8c9d5cf64c567
87 CLASSESP(mark | x,t)
Given an event in this context, what type is most likely?
A separate XGBoost multiclass mark model estimates the conditional probability distribution over all 87 modeled crime subtypes. It classifies event type; it does not decide whether an event occurs.
- OUTPUT
- conditional subtype distribution
- OBJECTIVE
- multi:softprob
- RUN ID
7efda77cdaec4a66a30321ea50b12ec8
03 / FEATURE SYSTEM
Static place context meets dynamic hourly state.
Both production models enforce the same 38-feature serving width: 25 national static features plus 13 values rebuilt or resolved for the inference hour.
Road and major-road density, intersections, dead ends, buildings, POI composition, road-type ratios, one-way roads, land-use entropy, and commercial-residential mix.
Population, median age and household income, poverty, unemployment, vacancy, renter occupancy, and household vehicle access.
Local hour and weekday plus cyclical encodings, resolved from each cell's geographic timezone for every inference timestamp.
Forecast temperature, relative humidity, and an explicit weather-availability signal for the selected hour.
Solar elevation, solar azimuth, daylight state, and categorical astronomical, nautical, civil, daylight, or night context.
04 / NATIONAL SPATIAL INFERENCE
One canonical model grid. Six rendering resolutions.
CrimeNet evaluates the intensity model once across the national H3-r9 domain. Lower resolutions are hierarchical sum reductions of that r9 field for serving and GPU map rendering—not coarser ML predictions.
r9r8r7r6r5r4
VIEWPORT-AWARE CRIMESENSE RENDERING · DETAIL FOLLOWS ZOOM AND CELL BUDGET
05 / ROLLING FORECAST ENGINE
Twenty-four independently inferred future states.
CrimeNet materializes a complete environmental feature snapshot and national intensity surface for every hour. Forecast frames are independently inferred model states, not interpolated map frames.
LIVE+6H+12H+18H+24H
GEOSPATIAL MODELING + INFERENCE SYSTEM
Powered by CrimeNet
CrimeSense is the product. CrimeNet is the data, feature, model, materialization, aggregation, and serving infrastructure underneath it.