Transcript Crime Risk Factors Analysis

Crime Risk Factors Analysis
Application of Bayesian Network
Introduction
• Urbanization causes the crime problem to increase in
magnitude and complexity
• To provide effective control, with limited resources, law
enforcement agencies need to be proactive in their
approach
• Since cities and crime pattern are dynamic in nature, a
system is required which can develop a model that
provide accurate predictions and must also constantly
updates changes various parameters over time
• BN has been adopted for inference and prediction in the
crime problem domain
Introduction Contd.
• The use of BN enable us to use the incomplete existing
data to setup initial model and to continue the
enhancement of the model’s predictive capabilities as
new data is added
Analysis of the Factors Affecting Crime
Risk
1. Identifying Crime Pattern Characteristic
2. Establish the relationship between various
factors
3. Determine the crime risk level
4. Recognize the crime data pattern (structure
learning)
5. Predict crime risk factors
Expert Probabilities Elicitations
• Eliciting probabilities from experts involve some
difficulties, therefore Keeney and von Winterfeldt’s
process for eliciting expert judgment was employed
• The formal process to elicit probabilities from expert
consists of seven steps
1.
2.
3.
4.
5.
6.
7.
Identification and selection of the issues
Identification and selection of the experts
Discussion and the refinement of the issues
Training for elicitation
Elicitation
Analysis, Aggregation and resolution of disagreement
Documentation and Communication
Data Processing
• The purpose is to prepare raw data for subsequent
process of analysis, it involves three steps
1. Data consolidation
2. Data Selection
3. Data Transformation
• A set of data containing 1000 records were obtained
• The data contain 20 variables, which were organized in
five groups of factors
• Population, Crime Locations, Types of Crimes, Traffic and
Environment
Bayesian Network Model
• The BN developed for this research project was based on
crime pattern analysis carried out by Brantingham and
Brantingham and the theory of crime control through
environmental design
• Pattern theory focuses on the environment and crime,
and maintains that crime location, characteristics of such
locations, movement path that brings offenders and
victims together at such locations, and people’s
perception of crime locations are significant objects for
study
Bayesian Network Model Contd.
• Pattern theory synthesizes Its attempts to explain how
changing spatial and temporal ecological structures
influence crime trends and patterns
• The model was constructed and tested using Hugin
software
• Conditional Probabilities Tables (CPT)
• Sample of CPT
– The probability of each input node was calculated using data
contained in the training example for each state, the variables
have five states, 0=very low, 1=low, 2=medium, 3=high and
4=very high
Learning Bayesian Network
1. Input
–
Consists of prior crime data & background knowledge
2. Output
–
Revised Bayesian Network
3. Learning Algorithm
–
–
To determine which links to be included in the DAG
To determine the parameters
A
Population
statistics, Crime,
Geographic Data
B
Learning
Algorithm
C
Revised Bayesian Network
The Expectation Maximization
Algorithm
• The Expectation Maximization (EM) algorithm
was used within the Hugin software for the data
• In the study since we have the complete set of
data, the EM algorithm was used to count the
frequencies of probabilities, it consist of two
iteration steps
– Expectation step (E-step) calculates the expectation of
the missing statistic
– Maximization step (M-step) maximizes a certain
function
• The two steps are performed until convergence
The Expectation Maximization
Algorithm
Initial network
and training
data
Computation
E-step
Expected
Count
Reparametrize
M-step
Updated
Network
The iteration steps of the Expectation Maximization Algorithm
Results
• The results shows that the factors that
considerably affected crime risk in Bangkok
Metropolitan area were environment, types of
crime, crime location, traffic and population.
• Crime risk probability given murder is “yes”
makes environment (drug sale and low
standard housing) as the primary reason on
the expected murder rate
Results
Factor
Name of variable
State very low
State very high
Environment
Drug Sale
0.2710
0.6935
Low Standard
Housing
0.0664
0.5795
Rape
0.1033
0.2762
Robbery
0.0983
0.4268
Nightclubs
0.0732
0.4548
Shopping center
0.2135
0.2071
Movie theater
0.2874
0.2259
Bank
0.2528
0.2538
Traffic
Traffic volume
0.1875
0.2797
Population
Pop Density
0.1734
0.2334
Crime
Location