Team Control Number
7071
Problem chosen
B
MLE & DRMA
Geographical Profiling Model
Abstract
Being familiar with the unsolved mystery of Jack the Ripper, serial crime has
proven to be a difficult puzzle for both criminal investigators and the police. However,
more emphases have been put on a suspect’s geographic patterns with the
development of Geographical Profiling Technique. Motivated by the Page-Rank
Algorithm used by Google search engine, we develop a Danger-Rank Matrix
Algorithm (DRMA) to rank the dangerous degree of the potential serial criminal
locations and predict the high-risk areas. The technique of Maximum Likelihood
Estimation (MLE) is used to examine the probability of crime in certain situation. The
model also can predict the location of next crime directly without considering where
the suspect lives. As a real world study, we successfully apply our model to the
notorious “black ghosts” case recently occurred in China. From the view of
implementation, our model is easy to be operated by the local police agency by
offering a criminal information database system.
Team # 7071 Page 1 of 17
Contents
1 Introduction ................................................................................................................. 2
1.1 Serial Crimes (Murders) ................................................................................... 2
1.2 Geographical Profiling Technique and Serial Crimes ...................................... 3
1.3 Objectives ......................................................................................................... 3
2 Simplifying Assumptions ............................................................................................ 3
About the serial criminals: ...................................................................................... 3
About the model:..................................................................................................... 4
3 The Model ................................................................................................................... 4
3.1 The Maximum Likelihood Estimation .............................................................. 4
3.2 The Danger-Rank Matrix Model ....................................................................... 6
4 Case Study .................................................................................................................. 8
4.1 The “Black Ghosts” Case in China ................................................................... 8
4.2 Peter Sutcliffe’s Case ...................................................................................... 11
5 Evaluating the Model ................................................................................................ 12
6 Conclusions ............................................................................................................... 13
7 Executive Summary for a Chief of Police ................................................................ 14
7.1 Overview of the Approach .............................................................................. 14
7.2 Situations Where the Approach is Appropriate ............................................... 15
7.3 Specific Recommendations for Implement ..................................................... 15
References .................................................................................................................... 16
Team # 7071 Page 2 of 17
1 Introduction
In 1888, the throats of five women prostitutes from the slums were cut prior to
abdominal mutilations in London by an unidentified killer lately named “Jack the
Ripper”. While not the first serial killer, Jack the Ripper's case was the first to create a
worldwide fear and frenzy of the people. The unsolved mystery of Jack the Ripper
still symbolizes the great difficulty to understand these dangerous predators. Serial
crime is a frightening and perplexing phenomenon which has proven to be a difficult
puzzle for both criminal investigators and the police.
However, the locations where the crimes happen are not completely random, but
instead have a degree of underlying structure. In order to provide guidance for the
police to deduce the likely whereabouts of the perpetrator of a series of linked
criminal offences, it is our task to develop a mathematical model to geographically
locate the suspect based on the past crimes.
1.1 Serial Crimes (Murders)
Serial murder is the intentional killing of two or more victims, at separate times and is
primarily predatory in nature. Before each murder, there is a building of tension
through fantasy, which is then relieved by the murder and the killer have a
“cooling-off period”. With an ultimate goal of murdering a victim, each victim
represents an intrinsic motive that fulfills a psychological need of the offender. (Bryan
Nelson, 2004)
Team # 7071 Page 3 of 17
1.2 Geographical Profiling Technique and Serial Crimes
A relatively recent development in the investigation is the emphasis on a suspect’s
geographic patterns: where a victim is selected, where the crime was committed or
where the bodies are dumped. This information can help the police to position the
potential area of the suspect’s residence. Take “the Yorkshire Ripper” Peter Sutcliffe
for example, the suspect was finally found living in the same town as the “center of
mass” technique suggested.
Geographic Profiling is defined as a methodology that uses the locations of a
connected series of crimes to determine the most probable area that an offender lives
in (Keith Harries, 1999). By referring to previous literatures, we can conclude that
crimes tend to occur at locations where suitable, in terms of profit and risk, offenders
encountered victims.
1.3 Objectives
Our main goal is to develop a mathematical model to provide guidance about possible
locations of the next crime (high-risk areas) based on the time and locations of the
past crimes scenes for the police agency. In order to achieve this goal, a “geographical
profile” of a suspected serial criminal is established. Other specific evidences such as
transportation and population density will be considered.
2 Simplifying Assumptions
About the serial criminals:
●That there is indeed a series of crimes that have been committed by the same
offender.
Team # 7071 Page 4 of 17
●That the offender is a 'local hunter' and not a 'poacher' who comes into the area to
commit crimes from elsewhere.
●That the same kind of people are the victims of the crime, since different behavioral
patterns are likely to be adopted by offenders depending on different targets.
● That the series of linked crimes are relatively complete and that any missing crimes
should not be "spatially biased". Police from different jurisdictions share information
completely with each other.
●That the perpetrator has a "single stable anchor point over the time period of the
crimes", which is to say that they haven't moved home recently or otherwise changed
their locations of crimes.
About the model:
●We assume that different kinds of serial crimes (usually including serial murders,
serial rapes and serial robberies) share the same characteristics. And we use the
features of serial murders to represent the characteristics.
●We assume that there is no gender difference between serial killers.
●We assume that the location where the crime is committed and where the bodies are
dumped overlaps.
3 The Model
3.1 The Maximum Likelihood Estimation
The probability of the commitment of serial crimes varies due to different features of
different locations. In order to estimate the possibility for a serial criminal to commit
crimes in block with certain feature(s), we divide the whole potential criminal
locations into different blocks.
Team # 7071 Page 5 of 17
Given that the event of the commitment of serial crime is A and the probability of
the commitment of serial crimes of any block lA is )( lAP . It should be noticed that
the value of )( lAP can usually be inferred by experience. Take three features: the
density of population, the transportation and the function of community into
consideration, we have:
1B : The event that the density of population is high.
1B : The event that the density of population is low
2B : The event that the transportation is well developed
2B : The event that the transportation is poorly developed
3B : The event that the function of community is perfect
3B : The event that the function of community is substandard
Figure 1 Block
●Suppose that Block 3 has high density of population, then the probability of the
commitment of serial crime can be estimated as
)(
)(
)|(
1
13
13
BP
BAP
BAP .
●Suppose that Block 4 has high density of population and well developed
transportation system as well, then the probability of the commitment of serial crime
Team # 7071 Page 6 of 17
can be estimated as
)(
)(
)|(
21
214
214
BBP
BBAP
BBAP .
●If the block has three or more than three features, the estimation method above still
works.
3.2 The Danger-Rank Matrix Algorithm
In order to narrow down the range of high-risk areas, we develop the Danger-Rank
Matrix Algorithm to rank the dangerous degree of potential criminal locations based
on their different features. Thus we need to label each location with “importance
score” based on the following hypotheses:
●We suppose that the offender commits crime at any location, for the other locations,
the more they are equipped with features which are convenient for the offender to
commit crime, the higher importance score the location will get.
Let us use Figure 2 to illustrate:
Figure 2
Team # 7071 Page 7 of 17
Suppose the offender may commit crime at any one of the four locations.
1. The arks in Figure 2 represent the travel direction of the offender, which means
the targeted location has features convenient for his crime and attract him to go
there.
2. We use jn to stand for the number of “attractive places”. For location 1, jn =2.
3. If one location has no attractive features for the serial criminal, meaning that the
location suffers no danger, we score 0 for that location.
●The more distinctive the feature of the area is, the higher the score will be. When
considering this standard, we can look for some kind of “signature behaviors”, which
can be seen as a reflection of the underlying traits of a specific offender. Because of
the outstanding distinctiveness of “signature behaviors”, we give high importance
score for areas fit them.
Given that nnijaA ][ , where
0
1
jij
na , if j links to i , then the formula is
mSAmM )1( where ]1,0(m
M : A weighted average of A and S
S : A nn matrix with all entries
n
1
A : Column-stochastic matrix
m : Constant
To rank the degree of danger of the four locations in Figure 2,
We have
1
312
32
3
3
4
421
3
41
2
4
1
x
x
xxx
x
xx
x
x
x
, and then
0100
3
1
01
2
1
3
1
00
2
1
3
1
000
A
We set 12.0m and the importance scores are 1288.01 x , 1855.02 x , 3488.03 x ,
Team # 7071 Page 8 of 17
3369.04 x .
Therefore, we can rank the locations by 2143 xxxx and location 1 is the
location with the highest risk.
Our scheme is based on Google’s Page-Rank algorithm, which ranks the importance
of web pages according to an eigenvector of a weighted link matrix. We analogize
looking for the high-risk area as searching information through search engine. Thus,
the features of certain areas such as well developed transportation system or high
density of rich people can be regarded as the “key words” when searching
information.
4 Case Study
4.1 The “Black Ghosts” Case in China
To examine the effectiveness and stability of our model, we apply our model to the
“black ghost” serial criminal happened in Anhui and Henan Province from April,
2002 to July, 2005.
We first estimate average serial criminal probability 018.0)( lAP according to
previous experience.
In this case, we find that the transportation is more important as the “black ghosts”
commit crimes along the highway, so we reasonably decide
that 12.0)|( 1 BAP l , 20.0)|( 2 BAP l , 17.0)|( 3 BAP l after several estimations.
According to the information above, we simply divide our potential locations into
four districts as follows:
Team # 7071 Page 9 of 17
Figure 3
Then we have
2
12
2
12
1
4
42
3
1
2
32
1
x
x
xx
x
x
x
xx
x
according to our Danger-Rank Matrix Model.
Next we assigned several values of m and choose the best one: 16.0m . At last we
get the ranking of the four locations and location 1 is the most dangerous one. In order
to ensure the stability of our model, we choose anther 165.0m and get the same
rank.
m
1x 2x 3x 4x
16.0m 0.3502 0.1871 0.2757 0.1871
165.0m 0.3497 0.1873 0.2758 0.1873
4231 xxxx
Team # 7071 Page 10 of 17
Figure 4 the testing of the case in China
Figure 5 the proving of the case in China
1
2
4
3
Team # 7071 Page 11 of 17
4.2 Peter Sutcliffe’s Case
We consider the time of committing the crime in Sutcliffe’s case from two aspects:
●The cycle of the criminal time: we use sin-function to imitate the cycle of the
criminal time by analyzing the “cooling-off” period between crimes.
Figure 6
●The day of week of serial crime: we use a radar diagram to analyze the day of week
of serial crime.
Team # 7071 Page 12 of 17
0
1
2
3
4
5
6
Monday
Tuesday
Wednsday
ThursdayFriday
Saturday
Sunday
Figure 7 Serial murder by day of week
It can be seen from the diagrams that the crime mainly happened on Saturday and a
longer “cooling-off” period comes after three short ones. When predicting the time of
next crime, the cycle and the day of week may be considered.
5 Evaluating the Model
The Danger-Rank Matrix Model has some obvious advantages. Firstly, it is
convenient for the police to operate. It works just like a database. Once the police type
in the key features, they can get the high-risk areas directly. Secondly, compared to
traditional technique, our Danger-Rank Matrix Model can predict the location of next
crime directly without considering where the suspect lives. This advantage guarantees
the “first-mover advantage” of the police. Thirdly, as can be seen from the case study
of the “black ghosts”, our model is relatively stable.
Team # 7071 Page 13 of 17
6 Conclusions
With the goal to generate a geographical profile and help the police to give prediction
of the location of next crime, we combine the Maximum Likelihood Estimation and
the Danger-Rank Matrix Algorithm (MLE & DRMM). Our model can help the police
to rank the degree of danger and narrow down the locations with high risk. We also
applied our model to the real case of the “black ghosts” in China to text the stability.
Team # 7071 Page 14 of 17
7 Executive Summary for a Chief of Police
7.1 Overview of the Approach
In order to foreclose further occurrence of serial crimes timely and effectively, we
develop a Danger-Rank Model to forecast the next potential criminal location
(“high-risk area). The procedural diagram below shows the steps of implementing the
Model:
Past serial criminal Record
Key features
Database
Potential
crime map
Importance scores
of Location
Prediction of possible
locations of next crime
Warnings to some
high-risk areas
New serial
criminal record
Analyze
Update
Information exchange Generate
Police Input
Rank
Output
Geographical Profile Model
Aid
Team # 7071 Page 15 of 17
7.2 Situations Where the Approach is Appropriate
Our approach is appropriate when the serial crimes fit the features as follows:
1) The offender is a “local hunter” and has a stable anchor point over the time period
of the crimes.
2) The victims of the crime are the same kind of people.
3) The information about the crime is abundant.
4) When the exact address of the suspect is unknown or hard to get, our model can
still predict the location of next crime.
7.3 Specific Recommendations for Implement
●In order to save the cost and police force, the police should concentrate on the
locations of the highest risk area.
●Always remember that geographical profiling only can offer some kind of support. It
should be combined with other detective techniques to enhance its efficiency. Other
factors such as the influence of the press and psychological profile should not be
ignored either.
Team # 7071 Page 16 of 17
References
[1] K.Bryan and T. Leise www.rose-hulman.edu/~bryan
[2] http://zmdgaj.zhumadian.gov.cn/Article/ShowArticle.asp?ArticleID=50
[3] http://en.wikipedia.org/wiki/Peter_Sutcliffe
[4] Bryan Nelson, 2004 http://www.deviantcrimes.com/defineserialmurder.htm
[5] Keith Harries, 1999 http://everything2.com/title/Geographic+profiling