Beruflich Dokumente
Kultur Dokumente
Neurocomputing
journal homepage: www.elsevier.com/locate/neucom
art ic l e i nf o
a b s t r a c t
Article history:
Received 12 June 2013
Received in revised form
20 October 2013
Accepted 16 February 2014
Communicated by Lixian Zhang
Available online 28 April 2014
This paper considers the solution to gang crime combined Analytical Hierarchy Process with Graph
Theory. The main purpose is to identify the conspirators and make a priority list based on the given
message trafc in a certain crime case. To identify one person, we rstly quantify the topics through
Analytic Hierarchy Process, then establish the network model based on Graph Theory and nally
conclude the relative relevance from this person to the known conspirators and non-conspirators. The
exibility of the model is illustrated and the results show that the method is effective.
& 2014 Elsevier B.V. All rights reserved.
Keywords:
Gang crime
Network
Graph Theory
AHP
Semantic analysis
1. Introduction
The knowledge explosion of science encourages the invention
of new technologies and contributes to the increase of economy
and enhancement of human lives. However, along with these
improvements, the crime also presents the high-tech and gang
characteristics [12]. For example, there are always some lawless
persons capturing loopholes and thus causing some troublesome
problems like internet intrusion and internet fraud [1,22], which
are detrimental to citizen's life and possessions. Compared to the
original types of criminal, the high-tech and gang characteristics
exaggerate both the detriments of these crimes and the complicity
to solve them. Although it is easy to capture some shadow conspiracy messages, it is a complex and time-consuming job to detect
most conspirators from the large message trafc.
In recent years, many algorithms and methodologies have been
developed in allusion to this kind of crime problem. Mohammad A.
Tayebi and Uwe Glsser applied mathematical models of crime data
and criminal activity as underlying semantic foundation to identify
organized crime structures in Co-offending Networks in [18]. In the
work of [10], N. Lchevin, C.A. Rabbath and P. Maupin proposed a
stability monitoring system of an Asset-Communications Network
Corresponding author.
E-mail addresses: dfgaoyulong@sina.com (Y. Gao),
xia_yuanqing@bit.edu.cn (Y. Xia), wu_suichao@126.com (S. Wu),
qjn83@163.com (J. Qiao).
http://dx.doi.org/10.1016/j.neucom.2014.02.041
0925-2312/& 2014 Elsevier B.V. All rights reserved.
122
initially adopted in this paper for solving the gang crime problem of
2012 The Interdisciplinary Contest in Modeling (ICM).
The Analytic Hierarchy Process is a structured technique for
organizing and analyzing complex decisions. Based on mathematics and psychology, it was developed by Thomas L. Saaty in the
1970s and has been extensively studied and rened since then. It
has been widely applied to decision making in various areas such
as economics, nance, politics, games and sports [28]. For example, Robert L. Nydick and Ronald Paul Hill used the AHP method to
structure the supplier selection process in [29]. Compared with
other methods like cosine transformation and circuit equivalence,
AHP is both exible and more smart, it can decrease the subjectivity of human judgment by classifying the hierarchy structure
in detail and normalization process, and the human judgment
can also gure out some latent useful messages to increase the
accuracy of the results. Graph Theory is the study of graphs, which
are mathematical structures used to model pairwise relations
between objects. In our model, the application of Graph Theory
can help decrease the subjectivity caused by human judgment
through meticulous mathematical calculation.
In order to nd out the conspirators and make a prior list, we
rstly quantify the topics through AHP and obtain their weights to
measure their suspicious degree. Then, based on Graph Theory, we
establish the network model and calculate the relevance between
any two persons using Dijkstra Algorithm [7,19,25]. Finally, we
take the sensitivity analysis and discuss the strengths and weaknesses of our model. Furthermore, we introduce keyword index
and semantic analysis to establish an Articial Intelligence System.
It can give the priority list of conspirators automatically and
diminish the subjectivity to enhance our model.
2. Problem formulation
The crime busting problem describes a gang crime. The conspirators and the possible suspected conspirators all work for the same
company in a large ofce complex. As shown in Fig. 1, ICM offers a
small set of messages from a group of 82 workers in the company that
they believe will help them nd the most likely candidates for the
unidentied co-conspirators and unknown leaders. Since the message
trafc is for all the ofce workers in the company, it is very likely that
some (maybe many) of the identied communicators in the message
trafc are not involved in the conspiracy. In fact, they are certain that
they know some people who are not in the conspiracy. Specic
information about the problem can be found in [1]. In this paper,
we focus on two requirements.
Requirement 1: It is known that the crime busting case has 83
nodes, 400 links (some involving more than 1 topic), over 21,000
words of message trafc, 15 topics (3 have been deemed to be
suspicious), 7 known conspirators, and 8 known non-conspirators.
Build a model and algorithm to prioritize the 83 nodes by
likelihood of being part of the conspiracy and explain the model
and metrics. Jerome, Delores, and Gretchen are the senior managers of the company. It would be very helpful to know if any of
them are involved in the conspiracy.
Requirement 2: How would the priority list change if new
information comes to light that Topic 1 is also connected to the
conspiracy and that Chris is one of the conspirators?
We specically focus on the topics given in the attachment and
model the crime busting using network analysis. Our goals are to
identify people in the ofce complex who are the most likely
conspirators;
123
has been extensively studied and rened since then [28]. The
procedure for using the AHP can be summarized as follows.
First of all, we decompose the complex problem into a hierarchy
of several more easily comprehended sub-problems, namely the
elements, each of which can be analyzed independently. Elements
of a hierarchy dominate those in the lower hierarchy and are
dominated by the upper hierarchy. The relationship of domination
from top to bottom forms the hierarchy structure.
As is shown in Fig. 2, we specically focus on the topics and establish the hierarchy structure of them. The hierarchy structure consists
of three layersTarget Layer, Criterion Layer and Program Layer.
In criterion layer, we classify the topics into 5 groups according to
their contents and context. Dene C1 : The Company Operation; C2 :
Spanish; C3 : The Lives of Employees and the Company's Environment;
C4 : Company's Safety Condition; C5 : Suspicious topics;
The program layer contains the 15 topics among which some
topics are subjected to more than 1 group. We dene Tk: Topic k
(k 1; 215).
After establishing the hierarchy structure, afliation between
the upper and the lower layer is also determined. We obtain the
judge matrices by comparing any two elements in a same level
based on 19 scale method.
For the criterion layers elements C1, C2, C3, C4, C5, which are
dominated by the target layersuspicious degree, based on pairwise comparisons of the elements, we obtain matrix U (the Judge
Matrix between the target layer and the criterion layer):
U ui;j 55
where uij represents the priority degree of criterion i to criterion j
under the target layer. We use the number 19 to represent the
priority degree, the larger the number is, the bigger the priority
degree is. Here the properties of U are (1) ui;j 4 0; (2) ui;j 1=uj;i ;
(3) ui;i 1. we obtain the judge matrices between the criterion
layer and the program layer in the similar way for obtaining the
judge matrix U. Note that the criterion layer has ve elements, we
should determine the judge matrix between each element, say C1,
and the topics it dominates in the program layer. For instance, we
dene Matrix U as the judge matrix between the criterion and
topics 1, 2, 10 and 14 in the program layer. We can also obtain U,
B2, B3, B4 and B5.
0
1
0
1
1 12 1 13 15
1 3 2 3
B
C
B2 1 2 1 1 C
B 1 1 1 1C
2
3C
B
B3
C
2
B
C
C
U B 1 12 1 13 15 CB1 B
B 1 2 1 2C
B
C
2
@
A
B3 2 3 1 1 C
1
@
2A
1 12 1
3
5 3 5 2 1
Table 1
Data of R.I.
Order
R.I.
Order
R.I.
1
0
9
1.46
2
0
10
1.49
3
0.52
11
1.52
B
B2 @ 1
1
1
B5
B
B4 B
B4
@
1
2
1
1
1
1
B
B
1
B2
1
B4
B
C
1 AB 3 B
B1
B
1
B
B1
@
1
1
5
1
4
1
2
1
6
1
1
5
4
0.89
12
1.54
1
2
5
1.12
13
1.56
1
4
1
2
1
2
1
2
1
2
1
4
1
4
1
4
0
1
6C
C
CB 5 B
1
@
5C
A
1
1
6
1.26
14
1.58
7
1.36
15
1.59
8
1.41
C
2C
C
4C
C
C
1C
C
C
1C
A
1
C
1A
max n
n1
C:I:
R:I:
124
where W
is the weight of the corresponding element on level 2
relative to the target layer and t 3 is the weight of the element on
level 3 relative to that on level 2.
As is shown in Table 4, we obtain the weights of the 15 topics in
the program layer relative to the target layer through calculation.
3.2.2. Establish the network model
With the weights of the topics known, we establish the singledirected weighted network model in order to determine the
relevance between any two persons.
As is shown in Fig. 1, the network graph contains 83 nodes and
400 links. The graph is represented by GV; E, where V is the set of
ofce personnel, E is the set of edges. From the basic denition of
adjacency matrix introduced in the general model, we obtain this
models adjacency matrix A8383 . And aij is an element of matrix A,
and indicates the adjacent relationship between node i and j.
If there is an edge from node i to node j, aij 1; on the contrary,
aij 0.
In the crime busting problem, we conclude a persons relevance
with the conspiracy by using the similar algorithm for calculating
the shortest path. And the larger the relevance is, the more
suspicious the person is.
We rstly determine the Input Matrix M 8383 . mij is one element
of M, and indicates the relationship between node i and j. If they are
adjacent, mij is the sum of the weights of all the topics between
We dene
8
1 if pi;j 0
>
<
if pi;j 1
di;j 0
>
:p
if pi;j a 1πj a 0
i;j
max
5.0153
Normalized eigenvector
ci
gi
Table 5
The specic classication.
Classication
Relevance
Character
Manager
0 o r i o 1:05
1:05 o r i o 1:15
1:15 o r i
Non-conspirator
Uncertain
Conspirator
Employer
0 o r i o 1:00
1:00 o r i o 1:15
1:15 o r i
Non-conspirator
Uncertain
Conspirator
Table 3
Test result for B1 B2 B3 B4 B5 .
C:R:
Table 2
Test result for U.
n
C.R.
these two adjacent nodes. If they are not adjacent, mij is innite.
Otherwise, the diagonal elements of M are 0. Dene a matrix D8383
whose elements indicate the relevance between nodes. In order to
obtain D8383 , we rstly determine the transition matrix P. pij is an
element of matrix.
(
mi;k mk;j if mi;k mk;j o mi;j
pi;j
5
mi;j
if mi;k mk;j Z mi;j
B1
B2
B3
B4
B5
0.0039
0.0173
0.0227
Table 4
The Topics' weight for Requirement 1.
Topic
Weight
1
0.0371
2
0.0610
3
0.0073
4
0.0154
5
0.0256
6
0.0357
7
0.1956
Topic
Weight
9
0.0077
10
0.0214
11
0.2732
12
0.0572
13
0.2272
14
0.0115
15
0.0164
8
0.0077
125
Table 6
The specic results with priority rank for Requirement 1.
Classication
Name(Number)
Conspirators
Elsie(7), Jean(18), Alex(21), Paul(43), Harvey(49), Ulf(54), Yao(67), Le(61), Lao(58), Hark(70), Cory(71), Bariol(75), Mai(62), Gerry(77), Cole(76),
Quan(63), Andra(72), Vind(52), Gretchen(4), Gretchen(32), Dayi(51), Phille(79), Beth(38), Olina(55), Hazel(8), Carina(73), Franklin(24), Douglas(40),
Elsie(37), Sandy(12), Marion(13), Karen(5)
Non-conspirators Chris(0), Paige(2), Darlene(48), Gard(74), Tran(64), Jia(65), Ellin(68), Este(78), Darol(59), Cha(56), Sheng(57), Seeni(81), Beth(14), Crystal(20),
Kristina(1), Stephanie(30), Dwight(28), Marcia(27), Chara(53), Han(69), Lois(45), Shelley(35), Delores(10), Erica(39), Kim(33), Julia(15), Claire(25),
Louis(46), Patricia(44), Jerome(34), Priscilla(36), Jerome(16)
Uncertain
Wesley(23), Sherri(3), Kristine(19), Marian(26), William(50), Patrick(6), Katherine(42), Eric(22), Wayne(29), Malcolm(9), Reni(82), Melia(66),
Neal(31), Fanti(80), Francis(11), Lars(60), Christina(47), Donald(41), Neal(17)
Note: The bold ones are known conspirators and non-conspirators. Grethen(4), Delores(10) and Jerome(16) are considered to be managers. The order of names of
conspirators is determined by the relevance ri from the largest to the smallest.
Table 7
The topics' weight for Requirement 1.
Topic
Weight
1
0.1467
2
0.0610
3
0.0073
4
0.0154
5
0.0256
6
0.0357
7
0.1591
Topic
Weight
9
0.0077
10
0.0214
11
0.2367
12
0.0572
13
0.1907
14
0.0115
15
0.0164
8
0.0077
Table 8
The specic results with priority rank for Requirement 1.
Classication
Name(Number)
Conspirators
Chris(0), Elsie(7), Jean(18), Alex(21), Paul(43),Harvey(49), Ulf(54), Yao(67), Le(61), Lao(58), Hark(70), Cory(71), Bariol(75), Gerry(77), Mai(62),
Andra(72), Cole(76), Vind(52), Quan(63), Dayi(51), Phille(79), Gretchen(4), Olina(55), Carina(73), Gretchen(32), Marion(13), Karen(5), Franklin(24),
Elsie(37), Hazel(8), William(50), Wesley(23), Sherri(3), Beth(38), Sandy(12), Kristine(19), Melia(66)
Non-conspirators Paige(2), Darlene(48), Tran(64), Jia(65), Ellin(68), Este(78), Gard(74), Darol(59), Cha(56), Sheng(57), Seeni(81), Crystal(20), Kristina(1), Dwight(28),
Chara(53), Beth(14), Stephanie(30), Han(69), Lois(45), Shelley(35), Marcia(27), Erica(39), Kim(33), Julia(15), Claire(25), Delores(10), Jerome(16)
Uncertain
Eric(22), Katherine(42), Lars(60), Patrick(6), Douglas(40), Malcolm(9), Marian(26), Wayne(29), Donald(41), Reni(82), Patricia(44), Jerome(34),
Christina(47), Fanti(80), Neal(31), Priscilla(36), Francis(11), Neal(17), Louis(46)
Note: The bold ones are known conspirators and non-conspirators. Grethen(4), Delores(10) and Jerome(16) are considered to be managers. The order of names of
conspirators is determined by the relevance ri from the largest to the smallest.
1
B1
B
B5 B
@1
1C
C
C
1A
and
T
w3
5 0:25; 0:25; 0:25; 0:25
126
Acknowledgments
References
4. Conclusions
We develop a network model to solve the crime busting
problem through several steps. We rstly apply AHP (Analytic
Hierarchy Process) to quantify the topics and obtain their weights.
[1] /http://www.comap.com/undergraduate/contests/mcm/contests/2012/pro
blems/S.
[2] W.J. Bao, Z.Y. Wu, The analytic hierarchy process method for integrate
distribution issue of agricultural products logistics, in: 2011 International
Conference of Information Technology, Computer Engineering and Management Sciences, 2011, pp. 342344.
[3] G. Cheng, Y. Li, J.Q. Chang, X.F. Sun, The research and application of vehicle
comfortableness based on analytic hierarchy process method, in: 2010 IEEE
17Th International Conference on Industrial Engineering and Engineering
Management, 2010, pp. 11241128.
[4] J.A. Danowski, Semantic network analysis of Islamist sources using time slices
as Nodes and semantic similarity as link strengths: some Implications for
propaganda analysis about Jihad, in: 2012 European Intelligence and Security
Informatics Conference, 2012, pp. 164171.
[5] M. Dasclu, C. Dobre, . Truan-Matu, V. CristeaBeyond, Traditional NLP: a
distributed solution for optimizing chat processing automatic chat assessment
using tagged latent semantic analysis. in: 2011 10th International Symposium
on Parallel and Distributed Computing, 2011, pp. 133138.
[6] J. Fei, H. Xu, Assessing computer network security with fuzzy analytic
hierarchy process, in: 2010 2nd International Conference on Advanced
Computer Control, vol. 5, 2010, pp. 204208.
[7] N. Jasika, N. Alispahic, A. Elma, K. Ilvana, L. Elma, N. Nosovic, Dijkstra's shortest
path algorithm serial and parallel execution performance analysis, in: 2012
Proceedings of the 35th International Convention, 2012, pp. 18111815.
[8] Y.J. Jin, A hybrid-strategy method combining semantic analysis with rulebased MT for patent machine translation, in: 2010 International Conference on
Natural Language Processing and Knowledge Engineering, 2010, pp. 14.
[9] M. Kudelka, V. Snasel, O. Lehecka, E. El-Qawasmeh, Semantic analysis of web
pages using web patterns, in: IEEE/WIC/ACM International Conference on Web
Intelligence, 2006, pp. 329333.
[10] N. Lchevin, C.A. Rabbath, P. Maupin, Toward a stability monitoring system of
an asset-communications network exposed to malicious attacks, in: 2011
American Control Conference, 2011, pp. 22102215.
[11] X.K. Li, Z.G. Zhu, Automatic object classication through semantic analysis, in:
2008 20th IEEE International Conference on Tools with Articial Intelligence,
2008, pp. 497504.
[12] J.Q. Liang, K. Zhang, F. Gao, Finding conspirators in criminal network, 2012
International Conference on Systems and Informatics, 2012, pp. 161166.
[13] Y.L. Liu, Y.X. Ouyang, Z. Xiong, IDR-SQE: combining semantic and statistical
relations to rene documents, in: 2009 International Conference on Information and Multimedia Technology, 2009, pp. 139143.
[14] V. Loia, M. Mattiucci, S. Senatore, M. Veniero, Computer crime investigation by
means of fuzzy semantic maps, in: 2009 IEEE/WIC/ACM International Joint
Conferences on Web Intelligence and Intelligent Agent Technologies, vol. 3,
2009, pp. 183186.
[15] V. Pinheiro, T. Pequeno, V. Furtado, D. Nogueira, Semantic inferentialist
analyser: a semantic analyser for natural language sentences, in: 2009 Seventh
Brazilian Symposium in Information and Human Language Technology, 2009,
pp. 8997.
[16] W. Song, S.C. Park, Analysis of web clustering based on genetic algorithm with
latent semantic indexing technology, in: 2007 Sixth International Conference
on Advanced Language Processing and Web Information Technology, 2007,
pp. 2126.
[17] E. Spyrou, G. Tolias, P. Mylonas, Y. Avrithis, A semantic multimedia analysis
approach utilizing a region thesaurus and LSA, in: 2008 Ninth International
Workshop on Image Analysis for Multimedia Interactive Services, 2008, pp. 811.
127