Deep Learning for Public Safety in Chicago and San Francisco

DEEP LEARNING FOR PUBLIC SAFETY: FIGHTING CRIME WITH OPEN CITY DATA

Alex Tellez, Michal Malohlava, and H2O.ai team

OPEN CITIES

Many major cities around the world

provide easily accessible public data sets

with years of historical data

Currently this data is underused

CHICAGO

OPEN CRIME DATA

Crime Dataset: Crimes from 2001 - Present Day~ 4.6 million crimes

THE WINDY CITY

Harvest Chicago Weather data since 2001

SOCIOECONOMIC FACTORS

Crimes segmented into Community Area IDsPercent of households below poverty, unemployed, etc.

SPARK + H2OWeather CrimesCensusWeatherWeather

Data munging

Spark SQL join

Deep Learning

Evaluate models

GOAL:For a given crime,

predict if an arrest is

more / less likely to be made!

LOAD DATA INTO H2O

Weather Data5k rows

Census Data78 rows

Crime Data~4.5 Mn rows

JOIN DATASETS

crimedata

weatherdata

censusdata

Using Spark, we join 3 datasets together to make one mega dataset!

CHICAGO VISUALIZATIONS

arrest rate season of crime

temperature during crime

community crime is

committed in

ARREST RATE BY TYPES OF CRIME

ARREST RATE VS % OF TOTAL CRIMES

Arrest Rate

% of all crimes recorded

A large proportion of crimes are thefts

Unfortunately, there is a much lower arrest rate for thefts than for less

prevalent crimes like gambling

SPLIT DATA INTO TEST/TRAIN SETS

training set arrest rate test set arrest rate

train model on this segment, 80% of data

validate the model on this segment (remaining 20%)

~40% of crimes lead to arrest

DEEP LEARNINGProblem:

For a given crime, is an arrest more / less likely?

Deep Learning:

A multi-layer feed-forward neural network that starts

w/ an input layer (crime + weather data)

followed by multiple layers of

non-linear transformations

DEEP LEARNING MODELDeep Neural Network w/ 2 layers of non-linear transformations

Binomial prediction: Is an arrest made? Yes/No

AUC on Training Data ~ 0.91!~ 3.5 Million Crimes

HOW’D WE DO?Train AUC ~ 0.91 Test AUC ~ 0.91

GEO-MAPPED PREDICTIONS

Because each of the crimes reported comes with latitude-longitude coordinates, we scored our hold out data using the trained model and plotted the predictions on a map of Chicago - specifically, the Downtown district.

SAN FRANCISCO

OPEN CITY, OPEN DATA

Crime Dataset: SFPD Incidents from 1/1/2003 - Present~1 Million Crimes

WEATHER ANYONE?

Harvest weather data from 1/1/2003

DATA INGESTION

Weather Data: Temp, Visibility, Precipitation, Cloud Cover

Crime Data: Category, Description,

Weekend, Arrest, etc

SF VISUALIZATIONSMost common crimes? When is crime happening most?

…midnight, noon, 6 PM

DEEP LEARNING MODELDeep Neural Network w/ 3 layers of non-linear transformationsTotal Run Time: 6 mins. 42 sec.

AUC ~ 0.95 on Training Data

VALIDATION TESTModel ‘trained’ on 80% of data, validated against remaining 20%

AUC = 0.95 on validation data

WHAT’S NEXT?Can deploy each model in real-time to increase public safety

and help police departments.

Map of Model Accuracy - For each point on the map (place of crime) we can have different colors based on model prediction (0.999 =

green, arrest likely vs. 0.67 = orange)

Run prediction for specific subsets of the data (i.e. most dangerous area)

We plan on doing all of the above!

Ensemble - Model average by running prediction models for Chicago + San Francisco which may increase accuracy more?