Team Control Number

12218
For office use only For office use only
T1 ________________ F1 ________________
T2 ________________ F2 ________________
T3 ________________ F3 ________________
Problem Chosen
T4 ________________ F4 ________________
C

2012 Mathematical Contest in Modeling (MCM) Summary Sheet

Type a summary of your results on this page. Do not include

the name of your school, advisor, or team members on this page.

Social Network Analysis in Crime Busting

Social network analysis (SNA) has been winning more and more attention in the last twenty
years. The method provides an insight into various networks. In this paper, we use SNA and
related techniques to analyze crime data and try to get potential criminal gang.
After first investigating the hidden features of a conspiratorial group, we consider one’s
closeness to the criminal gang as the main reason in deciding whether a staff member should be
regarded as a conspirator. The introduction of cooperation distance metric (CD-metric) combined
with the analysis on cooperation factor makes up our model’s major base. Within these definitions,
potential conspirators can be worked out by choosing those on the top of the ascending CD-metric
list. Highest values of CD-metric can be recognized as being the farthest away from the conspiracy,
in other words, they are the most innocent ones. A 12 members’ criminal group is dug out.
Before identifying the leading criminals, it is essential for us to quantify the ability of
people’s leadership. Centrality analysis suggests a good way in determining the focal point(s) in a
network. We make some amendments of the centrality measures so that it can be applied to a
directed graph. What’s more, we combine centrality measures(degree, closeness and betweenness)
and figure out the fusion values. We can form a ranking list of people’s ability of leadership then.
The leaders in the criminal group are sure to come to surface after comparing with the priority list
derived before.
In the refinement of the network model, we appeal to the theories of semantic network
analysis and text analysis. On the purpose of making a deeper exploration of identifying the
intrinsic character of one topic, we reapply the centrality analysis to data computing right after the
construction of a semantic web. Based on text analysis approach, we build up a vector space
model and compare the final outputs with ones obtained above, the results are robust.
This model’s application in other areas such as the biomedical domain has been discussed at
the end of this paper, excellent performance happens when accessed to enough amounts of data.
With a well-organized structure and a wide range of application, this methodology is sure to have
a bright future.
Keywords: Social Network Analysis, Centrality Measures, Semantic Web, Text Analysis

更多数学建模资料请关注微店店铺“数学建模学习交流”
https://k.weidian.com/RHO6PSpA
Team # 12218 Page 1 of 18

Contents

1 Introduction 2

2 Investigation Scenario Based on Crime Data Mining Model 2

2.1 General Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Explanation of the Model . . . . . . . . . . . . . . . . . . . . . . 3
2.4 Tests and Verifications . . . . . . . . . . . . . . . . . . . . . . . . 4

3 Analysis on the Current Case 6

3.1 Priority Lists of Potential Conspirators . . . . . . . . . . . . . . . 6
3.2 Performance with More Accurate Information . . . . . . . . . . . 7

4 Searching for Leaders 7

4.1 Centrality Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 8
4.2 Interpretation of Results . . . . . . . . . . . . . . . . . . . . . . . 10

5 Model Optimizing 11
5.1 Semantic Network Analysis . . . . . . . . . . . . . . . . . . . . . 11
5.1.1 Construction of Semantic Network . . . . . . . . . . . . . 12
5.1.2 Topic Selection . . . . . . . . . . . . . . . . . . . . . . . . 12
5.2 Text Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
5.2.1 Vector Space Model . . . . . . . . . . . . . . . . . . . . . 13
5.2.4 Corroboration of Model Reasonableness . . . . . . . . . . 14

6 Applications 15

7 Conclusion 15

1
Team # 12218 Page 2 of 18

1 Introduction
A crime is consist of a wide range of activities ranging from traffic violations
to organized fraud in financial field. The major challenge lying ahead of all the
authorities is how to accurately and efficiently analyze tremendous amount of
the crime data [1].
With great advance in the information technology, the financial crime hap-
pens more often than ever. Due to a German report published not long before,
the identified fraud cases related to card payments have risen 345% just from
2007 to 2009 [2]. A common analytical approach to this issue during recen-
t years is called “Social Network Analysis”(SNA), which can be viewed as an
innovation on mathematical perspective combined with sociology and crimi-
nology. This method can expose the hidden structural patterns in criminal
networks [3], which can be offered to the anti-crime departments in directing
them to necessary surveillance or interrogation scenarios, after several steps of
data processing.
According to Mena’s book, SNA is a data mining technique that reveals
the structure and content of a body of information by representing it as a set
of interconnected, linked objects or entities [4].
To further present the application of the SNA approach, we arrange paper
as followed:
i. In section 2, we propose a way to construct a model in identifying potential
conspirators with the help from SNA in the first place;
ii. In section 3, verified by the results of in small network, our model is applied
to a bigger one and we rank all the people according to their possibility of
being involved in the conspiracy;
iii. In section 4, based on the centrality analysis theory, we explore a way in
determining all the criminal leaders;
iv. In section 5, after referring to the powerful techniques–semantic network
analysis and text analysis, we make some amendments of our model by
considering the content and context of the message traffic;
v. At last, we discuss the model’s further developments when provided with
enormous amount of data and talk about the future application such as in
the biomedical area.

2 Investigation Scenario Based on Crime Data

Mining Model
2.1 General Assumptions
• There is only one criminal gang in the given network;
• Messages can not be classified emotionally from appearance;
• Occurrences of messages are irrelevant to time;
• Effects form data errors are negligible.

2
Team # 12218 Page 3 of 18

2.2 Definitions
Before proceeding to the real problem, we first make some definitions:
The multi-graph G = (V, E) is assumed to be a social network, where
V denotes the set of nodes and E denotes the set of edges. A person in the
network can be denoted by a node in the graph while edges stand for messages
with certain topics passing through different persons. Thus we can separate
these persons into several cliques, judging from the topics of communication.
Let G0 = (V 0 , E 0 ) be a supergraph corresponding to multi-graph G, where
V is the set of supernodes and E 0 is the set of superedges [5]. Each su-
0

pernode Pi0 represents topic i with associated people in multi-graph G and there
exists a superedge when supernodes connected by it possess common nodes.
The weight of one superedge is the number of the common nodes.

Figure 1: Multi-graph Figure 2: Supergraph

Figure 1 and 2 presented above display the transition from multi-graph to

the corresponding supergraph. Nodes in the multi-graph can be recognized as
persons and supernodes in the supergraph can be recognized as topics involved
with relevant people.
So the crime busting in a criminal network can be translated into: “Pro-
vided with a multi-graph G = (V, E) and a query list Q = {q1 , q2 , q3 , . . . , qi },
we aim to find out all the possible subsets of nodes associated with ones in Q
and form a list ordered by the measurements of association.”

2.3 Explanation of the Model

The researched crime network is connected by messages delivery, through
information transmitting, the whole conspiratorial network can function effi-
ciently. An common staff member may have no idea in knowing whom the
people really are, when having a conversation with them whereas the intrinsic
character of topic can determine whether a person is involved in the conspiracy
or not.
A piece of message is consist of countless information in many aspects.
It is impossible and unnecessary to deal with them all at the same time, let
alone composing these messages to a conclusion. In this case, we only focus our
attention on three main aspects: Topics, Direction and People to whom they
talked. However, the biggest challenge that we have confronted with is how to
combine these three kinds of information appropriately.

3
Team # 12218 Page 4 of 18

We proposed a structure of network here in which the nodes denote topics

of messages and edges denote people involved in them. Regarding the definition
listed above, these characters are all possessed by the supergraph. In the general
view, when a person switch his or her attention from one topic to another, we can
regard it as person “flows” among related topics, therefore the whole network
become alive.
Before the further exploration of this method, we should first analyze the
“similarity” among collected topics. We use “Cooperation Factor (CF)” [5] to
study the latent interrelationships.
Definition 2.3.1 (Cooperation Factor) In the supergraph G0 = (V 0 , E 0 ), the
Cooperation Factor between two supernodes Pi0 and Pj0 is:

2 × Nc (Pi0 , Pj0 )
CF (S 0 , Vi0 , Vj0 ) = , CF ∈ [0, 1]; (1)
Nt (Pi0 , Pj0 )

Where: Nc denotes the number of common nodes between Pi0 and Pj0 ; Nt denotes
the number of total nodes contained in the Pi0 and Pj0 .
It is obvious that not all conspirators can contact directly but they must
be very close for safety concern. To utilize this collective feature, it is essential
to define a Cooperation Distance (CD) [5]:
Definition 2.3.2 (Cooperation Distance Metric) In the supergraph G0 =
(V 0 , E 0 ) with a list of i person denoted by a query Q = {q1 , q2 , q3 , . . . , qi }. Let
v ∗ be a particular node in the multi-graph G, and V be the set of nodes in the
multi-graph G. Then, the Cooperation Distance of v ∗ is then defined as:
|Q|
ShortestP ath(S 0 , v ∗ , qi )
P
0 ∗ i=1
CD(S , v , V ) = ; (2)
|Q|

Where: the ShortestP ath(S 0 , v ∗ , qi ) is 0 when both qi and v ∗ belong to a same

supernode, otherwise is the value derived from shortest path algorithm of v ∗ and
qi with the weight of edges measured by CF.
The value of CF evaluates the level of correlation between two different
topics, then 1 − CF indicates the level of uncorrelation in the opposite. We pick
out the assumed conspiratorial topics and compare other unidentified topics with
it by accordingly assign different weights. Hence, we can apply the shortest
path algorithm to this theory and figure out the corresponding CDs of each
unidentified staff member.
Those people with lowest values can be considered as having the most
intimated communication with the criminal gang, which means they are more
likely to be the real members in a criminal gang.

2.4 Tests and Verifications

After computing all the data we can obtain from the EZ (an investigation
scenario provided by supervisor) case, concerning the points stated above, we
can make a list of the CDs’ value in the ascending order, which indicates the
possibility of a person being a conspirators in the opposite.

4
Team # 12218 Page 5 of 18

Name CD metric Rank

Dave 0.0000 1
George 0.0000 1
Ellen 0.0000 1
Carol 0.3333 4
Fred 0.3333 4
Harry 0.3333 4
Bob 0.7143 6
Inez 0.7143 6
Anne 1.0476 9
Jaye 1.4286 10

Table 1: Priority List for EZ Case

The top three in the list are the staff members with the highest possibilities
of being involved in the conspiracy derived from our analysis, where there are
already two are sure to be guilty. Moreover, Anne along with Jaye are both
identified to be the “farthest away” from the conspiracy. But it is sad to find
that Carol may still be misjudged as a conspirator and we may probably leave
Bob behind as well.

Figure 3: Discriminating Line for EZ Case

In order to determine a discriminating line in separating innocents from

conspirators, we calculate the difference among pairs of neighbors in the list
and pick out the biggest one through comparison. The hugest gap measured
by curvature between two staff members can be deemed as a boundary that
distinguishes two groups–the innocent one and the conspiratorial one for the
conspiratorial messages is sure to be restricted in a small group.
Although there still exists several demerits considering the results, the ex-
cellent structure of the investigation scenario encourages us to apply this method

5
Team # 12218 Page 6 of 18

to a larger criminal network with much more data.

3 Analysis on the Current Case

For now, the new network has 83 people involved, 400 collected messages,
over 21,000 words of messages traffic, 15 topics, 7 known conspirators, and 8
known non-conspirators. There are also several members with same names in
investigation which we just differentiate them by the denoted numbers. Based
on the SNA method we proposed before, we calculate each staff member’s CD
metric and arrange them in an ascending order list. Considering the list size,
we only list the top 40 to perform the results of our analysis.

3.1 Priority Lists of Potential Conspirators

We first work out the CF values of each adjacent vertices in order to quantify
the level of correlations among edges. Because there are three topics considered
to be involved in the conspiracy, we make a combination of these three initial
values into ones by setting equal weights and make an addition. Therefore CD
metrics can be determined by making the use of shortest path algorithm and
priority list may come to surface then.
The output of our analysis on the 3 topics which deemed to be conspiratorial
and 7 known conspirators are listed below:

Name CD metric Name CD metric

Elsie 0.0000 Paige 0.2859
Jean 0.0000 Neal 0.2859
Alex 0.0000 Priscilla 0.2859
Elsie 0.0000 Kristina 0.3641
Paul 0.0000 Sherri 0.3641
Harvey 0.0000 Franklin 0.3641
Ulf 0.0000 Marcia 0.3641
Cha 0.0000 Jerome 0.4224
Sheng 0.0000 Louis 0.4224
Darol 0.0000 Beth 0.4420
Yao 0.0000 Douglas 0.4420
Stephanie 0.2265 Crystal 0.4480
Kim 0.2265 Claire 0.4480
Beth 0.2265 Jerome 0.4480
Seeni 0.2265 Darlene 0.4480
Dolores 0.2340 Patricia 0.4525
Marion 0.2340 Jia 0.4525
Dwight 0.2340 Neal 0.4584
William 0.2340 Melia 0.4584
Lars 0.2340 Francis 0.4605

Table 2: Partial Priority List for Requirement 1

We note that the 11 persons on the top have the same values of CD metric,
on the other hand, indicates that they can be regarded as having the same

6
Team # 12218 Page 7 of 18

possibilities of being a conspirator. The 7 known conspirators being involved in

suspected group also demonstrates the reasonableness of our method.
Furthermore, we plot the discrimination line in the same way we displayed
in the small network by comparing the separation among adjacent persons in
the priority list. Thus, the possible conspirators are: Elsie, Jean, Alex, Elsie,

Figure 4: Discriminating Line for Requirement 1

Paul, Harvey, Ulf, Cha, Sheng, Darol and Yao.

3.2 Performance with More Accurate Information

The promising results obtained above motivate us to test our model’s per-
formance with more accurate information in the same network. When Chris
was identified to be a conspirator and Topic 1 was regarded as being associated
with conspiracy, we revaluate the level of correlations among total 15 topics.
After setting different similarities’ weights to particular edges accordingly, we
recompute these data and work out with a amended list and discriminating line.
The results are displayed in Table 3 and Figure 5.

4 Searching for Leaders

It’s hard to uncover all the leaders in a well-organized crime networks due
to its high complexity. But we can build a calibration system based on a theory
called Centrality Analysis [6,7] to identify some suspects who are more likely
to be the chief ones.

7
Team # 12218 Page 8 of 18

Name CD metric Name CD metric

Chris 0.0000 Beth 0.2877
Elsie 0.0000 Paige 0.3362
Jean 0.0000 Neal 0.3362
Alex 0.0000 Priscilla 0.3362
Elsie 0.0000 Crystal 0.3511
Paul 0.0000 Lois 0.3511
Harvey 0.0000 Ellin 0.3511
Ulf 0.0000 Han 0.3511
Cha 0.0000 Kristina 0.3525
Sheng 0.0000 Sherri 0.3525
Darol 0.0000 Franklin 0.3525
Yao 0.0000 Marcia 0.3525
Dolores 0.2869 Gretchen 0.3525
Marion 0.2869 Douglas 0.3525
Dwight 0.2869 Claire 0.4294
William 0.2869 Jerome 0.4294
Lars 0.2869 Louis 0.4294
Seeni 0.2869 Darlene 0.4294
Stephanie 0.2877 Jerome 0.4299
Kim 0.2877 Beth 0.4549

Table 3: Partial Priority List for Requirement 2

Figure 5: Discriminating Line for Requirement 2

4.1 Centrality Analysis

There are three measures of centrality that have been widely used in social
network analysis:

Betweenness Betweenness of one node is the number of geodesics (shortest

8
Team # 12218 Page 9 of 18

paths between any other two nodes in the network) passing through it:
gjk (i)
CB (i) = (3)
gjk

Where: gjk represents the number of shortest paths between any two
nodes and gjk (i) represents the number of shortest paths running through
node i;
Closeness Closeness is the sum of geodesics between the particular node and
every other node in the network:
N
X −1
CC (i) = d(i, j) (4)
j

Where: d(i, j) denotes the binary shortest distance between any two nodes
in the network;
Degree The original concept [6] of one node’s degree is its number of links as:
N
X
X
ki = CD (i) = xij (5)
j

Where: i denotes the focal node, j denotes other nodes in the network;
N is the node’s number and xij = {0, 1} are the entries in the adjacency
matrix.
However, some pioneers proposed a method which considering the weights
of each edge in a weighted network [8–10] by introducing a new measure
called strength:
XN
W
si = CD (i) = wij (6)
j

Where: i denotes the focal node, j denotes other nodes in the network;
N is the node’s number and wij are the entries in the adjacency matrix
whose value is positive when i is connected to j (so that it can reflect the
weight), otherwise is 0.
Now we utilize a method to combine both degree and strength with the
introduction of a turning parameter α [11]:
Wα si (1−α)
CD (i) = ki × ( )α = ki × sα
i (7)
ki
Where: α is a positive value which can be set according to different re-
quest. When the parameter is between 0 and 1, then it would be favorable
to obtain a high value, otherwise a low value would be preferable. In our
crime network analysis, we choose 0.5 to accomplish our goal.
The leadership of one individual conspirator can be “revealed” by one mea-
sure or the combination of two or three, whereas some extremely special cases
such as gatekeepers and receptionists are still of high possibility [12]. According
to the present research, this method can work out some results with fairly good
accuracy and is easily to spread out [13].

9
Team # 12218 Page 10 of 18

4.2 Interpretation of Results

It is obvious to see that we can form different lists according to different
measures, whereas we can hardly to tell which one is more important in deciding
one’s leadership.

Name Degree Closeness Betweeness

Chris 7.905 0.010 0.034
Kristina 8.743 0.011 0.045
Paige 14.477 0.013 0.063
Sherri 13.013 0.012 0.088
Gretchen 8.599 0.014 0.064
Karen 6.433 0.013 0.047
Patrick 7.497 0.012 0.042
Elsie 11.959 0.012 0.100
Hazel 9.000 0.011 0.064
Malcolm 6.559 0.011 0.052
Dolores 8.396 0.012 0.091
Francis 7.737 0.011 0.079
Sandy 8.927 0.012 0.058
Marion 7.254 0.013 0.086
Beth 8.546 0.012 0.044
Julia 12.956 0.013 0.100
Jerome 4.469 0.009 0.042
Neal 10.915 0.014 0.061
Jean 11.336 0.012 0.074
Kristine 8.975 0.012 0.085

Table 4: Partial Outputs of Centrality Analysis

Although these three measure possess excellent in measuring one’s ability

in leadership, some demerits still exist:

(i) Degree does not consider the overall structure of the network and will
inevitably lead to some misjudgments [14, 15];
(ii) The major challenge standing ahead of closeness is lack of capability to
deal with the disconnected components in networks [11];

(iii) Betweenness may function improperly when faced with some particular
cases such as a node has none shortest path running through it.
With so many unpredictable situations to concern, we simply place an
equal weight to all of them and get a fusion value as result. We choose 40 staff
members who possess the highest fusion value from our analysis, results are in
Table 5. They are deem to be ones who can have the most important impact on
the whole network, in other words, leaders in the company. Comparing it with
the priority list we work out, we can infer that Alex is a leader in the criminal
gang.

10
Team # 12218 Page 11 of 18

Name Fusion Value Rank Name Fusion Value Rank

Franklin 0.0763 1 Yao 0.3494 21
Julia 0.0803 2 Eric 0.3574 22
Alex 0.0884 3 Crystal 0.3695 23
Gretchen 0.0884 4 Wayne 0.3735 24
Darlene 0.0964 5 Francis 0.3775 25
Jerome 0.1406 6 Stephanie 0.3855 26
Elsie 0.1566 7 Hazel 0.3896 27
Paige 0.1606 8 Karen 0.3936 28
Sherri 0.1647 9 Christina 0.4016 29
Gretchen 0.1928 10 Beth 0.4096 30
Kristine 0.1968 11 Lois 0.4257 31
Neal 0.2008 12 William 0.4257 32
Dolores 0.2088 13 Beth 0.4297 33
Jean 0.2088 14 Kristina 0.4337 34
Marion 0.2570 15 Louis 0.4337 35
Donald 0.2610 16 Dwight 0.4498 36
Paul 0.2610 17 Harvey 0.4578 37
Patricia 0.2610 18 Elsie 0.4699 38
Neal 0.3173 19 Patrick 0.4779 39
Sandy 0.3414 20 Malcolm 0.4940 40

Table 5: Partial Leadership List on Fusion Values

5 Model Optimizing
To make our model function better, we need to take a look back on the
weight placed on each topics. In the previous work, we simply give the significant
topics an equal random initial value while others were determined by calculation
based on the quantified correlations among them. However, this process of
quantification can not provide a satisfied result precisely. To cross this barrier,
we switch our attention to more powerful techniques–Semantic Network Analysis
(SemNa) and Text Analysis.

5.1 Semantic Network Analysis

A semantic network is a network which represents semantic relations be-
tween concepts in patterns of interconnected nodes and arcs. It is often used
as a form of knowledge representation [16]. With the fast improvements of the
SNA approach, the application of this method has been extent to the Web [17],
encouraging us to apply it to the semantic network. In our current case, we need
to analyze the “crime ontology” as carefully as possible, where a multi-graph
can be constructed according to the theory of semantic network [18]. Here we
build a model based on the Semantic Network Analysis as the application
of SNA approach and centrality analysis we raised up before.

11
Team # 12218 Page 12 of 18

5.1.1 Construction of Semantic Network

In a semantic multi-graph, all nodes can be classified into two categories:
concepts and property. As they literally mean, one concept can be expressed
by using several properties, whereas different collections of properties can ex-
press different concepts. One property can be used to describe a wide range
of concepts while some of these concepts are specifical to this property. These
relations among concepts and property can be used to construct a semantic
network, which can be realized by the artificial intelligence system. Rules are
displayed below [18]:

Figure 6: Relations among Nodes

The relations exhibited above can be concluded as:

• There exists a directed edge from C1 to C2 , if concept C1 is a sub-concept
of C2 ;
• There exists a directed edge from P1 to P2 , if property P1 is a sub-property
of P2 ;
• A directed edge should be added from each major concept node to the
property node;
• Directed edges should be added from the property node to all range con-
cept nodes.

5.1.2 Topic Selection

In our crime busting scenario, we can regard the concepts and properties
defined before as topics and keywords separately. We assess the importance of
one topic in a criminal network by working out its centrality measures such as
degree, closeness and betweenness. One with higher value of the combination
of these three is supposed to be more crucial in a conspiracy. Thus, we can
place the weights for each topic according to its conspiratorial level based on
the content and context of messages’ transferring.
After revaluating the topics’ weights, we can reprocess the data to obtain
a better result that can be closer to the truth. The partial results with top 40
derived from semantic web have been exhibited in Table 6.

12
Team # 12218 Page 13 of 18

Name CD Metric Name CD Metric

Chris 0.0000 Lars 0.7437
Elsie 0.0000 Seeni 0.7437
Jean 0.0000 Paige 0.7445
Alex 0.0000 Neal 0.7445
Elsie 0.0000 Priscilla 0.7445
Paul 0.0000 Stephanie 0.7485
Harvey 0.0000 Kim 0.7485
Ulf 0.0000 Beth 0.7485
Cha 0.0000 Sherri 0.9474
Sheng 0.0000 Jerome 0.9474
Darol 0.0000 Louis 0.9474
Yao 0.0000 Neal 0.9506
Crystal 0.7423 Jerome 0.9506
Lois 0.7423 Douglas 0.9506
Ellin 0.7423 Melia 0.9506
Han 0.7423 Kristine 0.9597
Dolores 0.7437 Shelley 0.9597
Marion 0.7437 Donald 0.9597
Dwight 0.7437 Carina 0.9597
William 0.7437 Patrick 0.9770

Table 6: Partial Outputs from Semantic Web

5.2 Text Analysis

A large text databases potentially contain a great deal of knowledge, for text
itself is a complex collection comprised of enormous amount of information [19].
However, it cannot be analyzed by applying common data mining methods. In
order to extract valuable information from the messages, we construct a vector
space model to estimate the importance of various topics.

5.2.1 Vector Space Model

Before making a careful analysis on the importance of topics, we should first
build a bridge connecting words and topics’ similarity. Therefore we introduce
a measure called “Feature Item Weight” to quantify the impact of keywords on
a particular topic [20].
Definition 5.2.2 (Feature Item Weight) Feature item weight wik indicates
the importance of feature item Tk to the text Di :
tfik × log2 (N/dfk )
wik = qP  2 (8)
Tk ∈Di tfik × log2 (N/dfk )

Where: tfik is the time of the feature item’s (Tk ) appearance in text Di while
dfk is the number of text including feature item Tk . feature item’s (Tk ) ability
of differentiating the different kinds of text is weaker when the value of dfk is
higher; N denotes the total number of text, idfk = log2 (N/dfk ) is the frequency
of reverse text in which higher value indicate a stronger ability in differentiating
the different kinds of text.

13
Team # 12218 Page 14 of 18

Definition 5.2.3 (Text’s Similarity) Text’s similarity can be measured by

the cosine of included angle formed by the corresponding vectors [21]:
n
P
wij × wjk
k=1
Sim(di , dj ) = s (9)
n n
2 )( 2 )
P P
( wik wjk
k=1 k=1

Where: di = (wi1 , wi2 , wi3 , . . . , win ) is a vector representing text Di in this

vector space, and wik is the feature item weight.

5.2.4 Corroboration of Model Reasonableness

The value of the Sim can be used in weights’ determination through a more
accurate way. When reapply the method to the big network in current case, we
obtain a priority list resembled to the former one in Table 3, where the member
of the criminal gang are still the same while some slight modification of the
CD metrics of the innocent ones. Nevertheless, we are still quite confident in
a better performance of this model when accessed to a richer database. It is

Name CD Metric Name CD Metric

Chris 0.0000 Dolores 0.9271
Elsie 0.0000 Marion 0.9271
Jean 0.0000 Beth 0.9271
Alex 0.0000 Julia 0.9271
Elsie 0.0000 Jerome 0.9271
Paul 0.0000 Neal 0.9271
Harvey 0.0000 Franklin 0.9271
Ulf 0.0000 Claire 0.9271
Cha 0.0000 Marcia 0.9271
Sheng 0.0000 Dwight 0.9271
Darol 0.0000 Stephanie 0.9271
Yao 0.0000 Gretchen 0.9271
Crystal 0.9174 Kim 0.9271
Lois 0.9174 Jerome 0.9271
Ellin 0.9174 Shelley 0.9271
Han 0.9174 Priscilla 0.9271
Kristina 0.9271 Beth 0.9271
Paige 0.9271 Douglas 0.9271
Sherri 0.9271 Patricia 0.9271
Patrick 0.9271 Louis 0.9271

Table 7: Partial Outputs of Vector Space Model

amazing that the members in criminal group are unchanged, where reversely
corroborate the reasonableness of the network model.

14
Team # 12218 Page 15 of 18

6 Applications
Our method in analyzing the network with a complex database is comprised
of social network analysis, centrality analysis, semantic network analysis and
text analysis. Each of them can easily be extent to different areas with the same
intrinsic mechanism as we utilized, and some present researches even explain the
reasons why these methods could possess a wide range of application with social
concerns [22].
For now, text mining has become a prevalent automated method in exploit-
ing the tremendous amount of knowledge in the biomedical literature such as
the most common ones: rule-based or knowledgebased approaches, and statis-
tical or machine-learning-based approaches [23]. The former one mainly focus
on analyzing the structure of messages with sorts of knowledge while the latter
one stresses on classifying full sentences. Several real performances of the se-
mantic web application have already been surveyed in various biomedical area
with knowledge integration and exploration [24].
With the useful information extracted from the messages transferred in
the crowds, we can appeal to the semantic network analysis and text analysis
for a deeper insight. The theory of social network analysis provides a solid
base to construct a semantic web in which content and context of one piece of
message are under much more concern. Within a social group, the difference
between each individual lies in the social experience and their relation with
others, and these difference would eventually lead to a division of the social
work, bringing various networks into function. Centrality can provide a good
measurement of the structural importance of one node in the whole network [25].
This method of importance quantification can help us with search for the focal
point(s). Therefore, we can be directed to a efficient analyze by combining
these methodology into one as a whole. A massive amount of data can enrich
our sample collection and a good artificial intelligence system can enhance our
semantic network analysis’s ability in information categorizing which strengthen
the reasonableness of the whole model.

7 Conclusion
It’s not an easy task to formulate a well-organized crime busting scenario.
In order to accomplish our goal, we first discussed about the features one crim-
inal gang could have. Based on the principle of secretness in a conspiracy, we
defined a cooperation distance (CD metric) to decide whether a person should
be regarded as being involved in the conspiracy for we thought one’s closeness
with the criminal group is a fairly conclusive argument.
Centrality measures provided a good way for us in detecting the leaders in
the criminal network. Some modifications were proposed by us in order to apply
the theory in a directed graph. Through comparison with the given information,
it was delightful for us to find that the true mangers possess high grades in the
list of leadership measurement.
In the model’s refinement work, we appealed to the techniques of semantic
network analysis and text analysis. Visualization may be the most fundamental
difference between the traditional analysis and the network research in nowadays
study. After making a development of the Social Network Analysis, we success-

15
Team # 12218 Page 16 of 18

fully constructed a semantic web of topics and keywords connected by linguistic

association. Although objectiveness of this analytical method is fascinating,
the unpredictable abilities of data processing for different artificial intelligence
system is a major barrier impeding the development of the semantic network
analysis. We hope this situation can get a better changed with corresponding
remediation.
A non-prominent output later obtained from the addition of text analysis
is believed to result from being in lack of enough message simples. But the final
results was almost in acceptance because of the high resemblance to former
ones. According to the statistical principles, we are expected to have a better
consequence with more accuracy. The promising results persuade us to put
more faith in a good application of this model before long.

16
Team # 12218 Page 17 of 18

References
[1] H. Chen, W. Chung, J.J. Xu, G. Wang, Y. Qin, and M. Chau. Crime data
mining: a general framework and some examples. Computer, 37(4):50–56,
2004.
[2] B. Wiesbaden. Polizeiliche kriminalstatistik. Bundesrepublik Deutschland.
Berichtsjahr, 2007.
[3] M.K. Sparrow. The application of network analysis to criminal intelligence:
An assessment of the prospects. Social networks, 13(3):251–274, 1991.

[4] J. Mena. Investigative data mining for security and criminal detection.
Butterworth-Heinemann, 2003.
[5] A.M. Fard and M. Ester. Collaborative mining in multiple social networks
data for criminal group discovery. In Computational Science and Engineer-
ing, 2009. CSE’09. International Conference on, volume 4, pages 582–587.
IEEE, 2009.
[6] L.C. Freeman. Centrality in social networks conceptual clarification. Social
networks, 1(3):215–239, 1979.
[7] J. Xu and H. Chen. Criminal network analysis and visualization. Commu-
nications of the ACM, 48(6):100–107, 2005.
[8] A. Barrat, M. Barthelemy, R. Pastor-Satorras, and A. Vespignani. The
architecture of complex weighted networks. Proceedings of the National
Academy of Sciences of the United States of America, 101(11):3747, 2004.

[9] M.E.J. Newman. Analysis of weighted networks. Physical Review E,

70(5):056131, 2004.
[10] T. Opsahl, V. Colizza, P. Panzarasa, and J.J. Ramasco. Prominence
and control: The weighted rich-club effect. Physical review letters,
101(16):168702, 2008.

[11] T. Opsahl, F. Agneessens, and J. Skvoretz. Node centrality in weight-

ed networks: Generalizing degree and shortest paths. Social Networks,
32(3):245–251, 2010.
[12] W.E. Baker and R.R. Faulkner. The social organization of conspiracy:
Illegal networks in the heavy electrical equipment industry. American So-
ciological Review, pages 837–860, 1993.
[13] V.E. Krebs. Mapping networks of terrorist cells. Connections, 24(3):43–52,
2002.
[14] D.J. Brass. Being in the right place: A structural analysis of individual
influence in an organization. Administrative Science Quarterly, pages 518–
539, 1984.
[15] S.P. Borgatti. Centrality and network flow. Social Networks, 27(1):55–71,
2005.

17
Team # 12218 Page 18 of 18

[16] J.F. Sowa. Semantic networks. Encyclopedia of Cognitive Science, 1992.

[17] J.M. Kleinberg. Authoritative sources in a hyperlinked environment. Jour-
nal of the ACM (JACM), 46(5):604–632, 1999.
[18] B. Hoser, A. Hotho, R. Jäschke, C. Schmitz, and G. Stumme. Semantic
network analysis of ontologies. The Semantic Web: Research and Applica-
tions, pages 514–529, 2006.
[19] T. Nasukawa and T. Nagano. Text analysis and knowledge mining system.
IBM Systems Journal, 40(4):967–984, 2001.
[20] Qi Shiwei. Web text mining and its application in correlation analysis
between events (in chinese). Master’s thesis, National University of Defense
Technology, 2008.

[21] Shi Zhiwei, Liu Tao, and Wu Gongyi. An effective and efficient algorithm for
text categorization (in chinese). Computer Engineering and Applications,
41(29):180–183, 2005.
[22] A. Sih, S.F. Hanser, and K.A. McHugh. Social network theory: new insights
and issues for behavioral ecologists. Behavioral Ecology and Sociobiology,
63(7):975–988, 2009.
[23] K.B. Cohen and L. Hunter. Getting started in text mining. PLoS compu-
tational biology, 4(1):e20, 2008.
[24] H. Chen, L. Ding, Z. Wu, T. Yu, L. Dhanapalan, and J.Y. Chen. Se-
mantic web for integrated network analysis in biomedicine. Briefings in
bioinformatics, 10(2):177–192, 2009.
[25] K. Chan and J. Liebowitz. The synergy of social network analysis and
knowledge mapping: a case study. International journal of management
and decision making, 7(1):19–35, 2006.

18