ArticlePDF Available

SoNA: A Knowledge-based Social Network Analysis Framework for Predictive Policing

Authors:

Abstract and Figures

Major incidents can disturb the state of balance of a society and it is important to increase the resilience of the society against such disturbances. There are different causes for major incidents, one of which are groups of individuals, for example at demonstrations. The ideal way to handle such events would be to prevent them, or at least provide information to ensure the appropriate security services are prepared. Nowadays, a lot of communication, even criminal, takes place in social networks, which, hence, provide the ideal ground to gain the necessary information, by monitoring such groups. In the present paper, we propose an application framework for knowledge-based social network monitoring. The ultimate goal is the prediction of shortterm activities, as well as the long-term development of potentially dangerous groups, based on sentiment and topic analysis and the identification of opinion-leaders. Here, we present the first steps to reach this goal, which include the assessment of the risk for a major incident caused by a group of individuals based on the sentiment in the social network groups and the topics discussed.
Content may be subject to copyright.
147
International Journal
o
n Advances in Intell
igent Systems, vol
10
no
3
&
4
1
7
, http://www.iariajournals.org/intelligent_sys
tems/
201
7
, © Copyright by authors, Published under agreement with IARIA
-
www.iaria.org
SoNA: A Knowledge-based Social Network Analysis
Framework for Predictive Policing
Michael Spranger, Hanna Siewerts, Joshua Hampl, Florian Heinkeand Dirk Labudde∗†
University of Applied Sciences Mittweida
Faculty Applied Computer Sciences & Biosciences
Mittweida, Germany
Email: {name.surname}@hs-mittweida.de
Fraunhofer
Cyber Security
Darmstadt, Germany
Email: labudde@hs-mittweida.de
Abstract—Major incidents can disturb the state of balance of a
society and it is important to increase the resilience of the society
against such disturbances. There are different causes for major
incidents, one of which are groups of individuals, for example
at demonstrations. The ideal way to handle such events would
be to prevent them, or at least provide information to ensure
the appropriate security services are prepared. Nowadays, a lot
of communication, even criminal, takes place in social networks,
which, hence, provide the ideal ground to gain the necessary
information, by monitoring such groups. In the present paper,
we propose an application framework for knowledge-based social
network monitoring. The ultimate goal is the prediction of short-
term activities, as well as the long-term development of potentially
dangerous groups, based on sentiment and topic analysis and the
identification of opinion-leaders. Here, we present the first steps
to reach this goal, which include the assessment of the risk for
a major incident caused by a group of individuals based on the
sentiment in the social network groups and the topics discussed.
Keywordsforensic; opinion-leader; topic mining; expert system;
text analysis; classification; sentiment analysis
I. INTRODUCTION
The representation and communication of individuals, com-
panies and organizations, using the Internet, especially social
networks, has become the standard in our society. Even though
social networks are successful and have progressed throughout
these past years, they have also contributed to the formation of
new criminal energy. As already mentioned in [1], in particular,
the provision of an infrastructure for rapid communication
and the possibility to exchange ideas, pictures etc. in private
and protected environments, which are difficult to control by
investigators - if at all - enables radical or extreme political
groups, criminal gangs or terrorist organizations to use Social
Networks as a tool to plan, appoint and execute criminal
offenses. These groups often use large-scale events with a
high degree of group dynamics to promote their ideas. Events,
such as sporting events, demonstrations or festivals, cause high
expenses on security personnel. The inherent group dynamics
cause a great uncertainty and unpredictability concerning the
development of such events and make it difficult to estimate
how much security personnel is needed. For example, in
2014 the police officers spent more than two million working
hours just on securing soccer games in Germany [2]. Tar-
geted and automated monitoring of social networks, taking
into account the applicable legal provisions, can particularly
support strategic security planning as well as the development
of effective prevention strategies. As a positive side effect,
the subjective sense of security of the users is strengthened.
Authorities of the federal office for the protection of the
constitution, as well as intelligence services, are aware of
the importance of social networks as a source for important
information and increasingly focus on extracting and analyzing
this information. However, at this point the extraction and
evaluation of the information is done manually. Taking into
account the increasing number of users worldwide – currently,
for example, approximately 40 % of the population worldwide
uses social networks – it has to be noticed that there is an
enormous amount of potential profiles or communication to
be monitored. This demonstrates the need for an automated
solution that is capable of handling this amount of data and
the resulting complexity.
Consequently, the design of an application framework,
namely Social Network Analyzer (SoNA), for monitoring
groups and organizations in Social Networks as key elements
of critical events is presented to assist decision-makers. A
prototype implementing parts of this framework for monitoring
publicly accessible Facebook data is discussed.
The paper is segmented in six sections. The first two
sections following the introduction discuss the concept of
predictive policing as well as give a short overview about the
language characteristics in Social Networks. These sections
are followed by an outline of the framework, which is still
under development, including how the dangerous militant
profiles can be selected, how the risk of an event can be
assessed and the opinion-leaders can be identified. In Section
V, a prototypical implementation including its architecture and
currently available features is presented. Finally, the paper ends
with a conclusion, also discussing the progress of the work and
its future development.
II. PREDICTIVE POLICING AS A TO OL F OR RESILIENCE
ENGINEERING
A major incident includes a great number of casualties
and/or severe property damage [3]. At large-scale events, such
148
International Journal
o
n Advances in Intell
igent Systems, vol
10
no
3
&
4
1
7
, http://www.iariajournals.org/intelligent_sys
tems/
201
7
, © Copyright by authors, Published under agreement with IARIA
-
www.iaria.org
as described above, there is always a possibility for a major
catastrophic event to happen. However, whether or not it will
happen is usually difficult to predict. Resilience is the ability
of a socio-ecological system to recover from disturbances, for
example a major catastrophic event, and retain or regain its
identity, functions, structures and its ability to respond [4]. In
a study about resilience the German Academy of Technical
Sciences (acatech) developed a resilience cycle based on the
Social Resilience Cycle by Edwards [5], which includes the
following five stages: prepare, prevent, protect, respond and
recover [6]. In order to return quickly to the defined secure
state of balance [7] it becomes necessary to apply resilience
engineering [8] in the sense of a technical support system,
which allows to anticipate the disaster situation [6]. Crime
that arises from dynamic groups at large-scale events as well
as organized and especially political motivated crime regularly
disturb the state of balance. Information gained from moni-
toring activities of such groups in the Internet and especially
Social Networks can be used to predict the probability of
such catastrophic events beforehand. Accordingly, the National
Institute of Justice in the USA defined Predictive Policing as
follows:
“Predictive policing, in essence, is taking data
from disparate sources, analyzing them and then
using the results to anticipate, prevent and respond
more effectively to future crime.” [9]
The knowledge gained from the monitoring of suspicious
groups in Social Networks directly contributes to an increase in
resilience in the stages Prepare and Prevent of the resilience
cycle [5] [10]. Therefore, the development of an automated
solution to monitor Social Networks is an important step of
resilience engineering.
III. CHARACTERISTICS OF SOCIAL MEDI A LAN GUAGE
While the language used in chat rooms is one of the most
researched topics [11], language used on the social media
site Facebook seems to be one of the least researched, which
is evident in the small amount of literature covering that
topic [12]. Zappavigna [12] suggests that one reason might
be the combination of several genres on one social media
site, making the analysis very complex. Even though it is
impossible to generalize the language found on the Internet
[11], studies about language use for example in chat rooms
or on microblogging sites, combined with the scarce literature
covering some linguistic aspects on Facebook gave a starting
point for an analysis. The focus of this paper is on posts and
comments and, therefore, excludes messages written on the
instant messenger.
In order to get a first impression of the language used
in Facebook groups, a small corpus was created using posts
and comments from different Facebook groups, relevant to
the application of SoNA (see Table I). The structure of a
“conversation” in a Facebook group is very different to the
structure for example of a chat conversation. The starting point
for a “conversation” on a group wall on Facebook is always a
post, often written by the group itself. Afterwards, users can
write a comment or reply to an already existing comment. In
comparison to a chat conversation the user is not expected to
write a comment immediately after a post was posted or write
a reply to a comment from another user. In fact, they do not
have to reply at all. This leads to the fact that “conversations”
in Facebook groups are not almost-synchronous as in a chat
conversation, yet clearly asynchronous [13] [14]. Therefore, it
might be questioned whether to talk about “conversations” at
all. Nonetheless, whenever users start a discussion on a group
wall and reply to each other’s comments within minutes, these
conversations look very similar to chat messages. Overall,
this “conversation” structure on Facebook leads to a highly
complex way of communication, which makes the analysis of
the language used and the meaning created difficult.
Furthermore, the wall on Facebook allows the users to
include multimodal communication, by posting pictures or
videos, either with a comment or with words included for
example in the picture. Additionally, often posts include ref-
erences to other websites or users simply repost a post from
someone else. Another aspect that makes the automated anal-
ysis of meaning difficult is the language itself. Characteristics
taken from studies on other Internet-based communication
were used as features in an annotation with the UAM corpus
tool of the small corpus mentioned above [11] [15] [16]
[17] [18]. The results show clearly that there seems to be a
difference between posts and comments. For example, orality,
especially colloquial language, typing errors and lower case
spelling of nouns seem to be more common in comments. In
comparison, hashtags seemed to be used more often in posts
than in comments. Furthermore, comments and posts can be
distinguished by their length. While the length of posts varies
between zero words (e. g., pictures) up to 892 words, the length
of comments varies from 1 word up to 92 words. Moreover,
these numbers show that in comments one can often find
incomplete sentences. Even though, it seems that the typical
features found in chat conversations are not used as often in
comments and posts on Facebook, they are still present and
create a challenge for the automated analysis used in SoNA.
Especially, emoticons make the analysis of meaning difficult,
because the way in which they are used to create meaning is
complex and they can also be used to create irony [19]. This
is why, so far, the sentiment analysis used in SoNA is based
on word and not sentence level.
TABLE I: Summary of the corpus created under this work
including different types of Facebook groups.
subcorpus posts subcorpus comments
type # groups posts words comments words
right-winged 5 46 4539 97 1559
left-winged 5 48 5003 94 1618
soccer ultras 2 20 1211 40 323
total 12 114 10753 231 3500
IV. OUT LI NE O F A FRA ME WO RK
The analysis of social networks from the point of view of
security policy pursues two main objectives. The first one is
the identification and estimation of potential dangers, including
their scope and location. The second one is to enable security
forces to plan in the long-term. In order to do so, it is of
special interest how a group is developing in terms of their
size growth, their orientation or radicalization and the increase
149
International Journal
o
n Advances in Intell
igent Systems, vol
10
no
3
&
4
1
7
, http://www.iariajournals.org/intelligent_sys
tems/
201
7
, © Copyright by authors, Published under agreement with IARIA
-
www.iaria.org
Topic
Analysis
𝜗 ∈ 𝜗𝑟𝑖𝑠𝑘
Sentiment
Analysis
𝑆 > 𝜀
NER
Location
NER
Time
Geo-
coding
Knowledge
Profile
Selection
𝑃𝐶
𝑓𝑟𝑖𝑠𝑘(𝜗, 𝑆, |𝑃|)
𝜃𝑟𝑖𝑠𝑘
Long-Term Development Prediction
Long-Term Risk
Profiles
Associated
|𝑃𝐴|
Leader
Multiplier
𝑃𝐿, 𝑃𝑀
Risk Assessment
Visualization
Figure 1: The proposed process chain for monitoring social networks.
in their propensity to violence. This section discusses basic
concepts of a framework that addresses these tasks.
The proposed framework allows decision-makers of secu-
rity forces, for example in the police’s management and control
centers, to identify and predict areas with high levels of crime.
As a result, it is possible to deploy forces more efficiently
depending on the specific situation. Thus, if, for example
during a debate about the policy regarding refugees on publicly
accessible pages of a social network, users loudly advocate
arson attacks on refugee homes, decision-makers can now put
security forces and specialized investigators on standby. If, on
the other hand, before a soccer games, violent fans or fans in
general do not seem to plan any riots, it may be sufficient to
return to the minimum number of necessary staff to secure the
event.
Another goal is to predict the long-term development of
potentially violent groups. Such a prediction may include state-
ments about the expected development of their membership,
but also evidence of a possible increase in radicalization in the
future. With this information, executives will be able to plan
resources and make infrastructural decisions in the long term.
If, for example, a district becomes, in the future, a point of
attack for various, growing and violent political groups, due
to certain circumstances, this information could lead to the
construction of an additional police station or the expansion of
the forces of an existing one. The development of a framework
for the automated analysis of data from social networks with
the aim of more effective crime prevention and defense in the
long and short term, makes it an application from the field of
predictive policing as defined in Section II.
In particular, the following tasks must be addressed by the
framework:
1) selection of potential profiles of dangerous militants,
2) assessment of the probability that the danger occurs,
3) determination of location and time of risks.
In order to meet the special needs and challenges of
forensics, especially with regard to the dynamics of language in
social networks, it is necessary to resort to expert knowledge.
This knowledge can be represented in the form of a Forensic
Topic Map (FoTM) as explained in detail in [20]. In particular,
abstract threats are modeled here, which form the basis for
the assessment and evaluation of the communication content.
Figure 1 shows the entire process chain for the proposed frame-
work. All process steps except for the long-term prediction,
which will be covered in future work, are explained in more
detail in the following subsections.
A. Selection of dangerous militants profiles
The selection of so-called dangerous militants profiles
ensures that profiles are not arbitrary selected and is thus es-
sential to the observance of data privacy protection regulations.
Furthermore, it focuses the monitoring on those profiles and
thus regulates the limitation of the analysis effort. Even though
the monitoring is limited to public profiles, and therefore all
information publicly available, it is important not to violate
the individual feeling of freedom, especially the freedom of
speech as regulated in the legal framework of the respective
legislature. The concept of the potential attacker was defined
by a German working group, consisting of the heads from
150
International Journal
o
n Advances in Intell
igent Systems, vol
10
no
3
&
4
1
7
, http://www.iariajournals.org/intelligent_sys
tems/
201
7
, © Copyright by authors, Published under agreement with IARIA
-
www.iaria.org
the State Offices of criminal investigations and the Federal
Criminal Police Office, for the scope of German law as
follows:
A dangerous militant is a person in whom
certain facts justify the assumption that they will
commit politically motivated offenses of consider-
able importance...” [21, translated by H.S.]
The extent to which this definition can be extended to
other areas of organized crime and gangs, without a political
motivation, remains to be legally clarified. Based on that
concept, a dangerous militant profile can now be defined as
follows:
A dangerous militant profile is the profile of a
dangerous militant in a social network. All profiles
associated with this profile are part of the extended
dangerous militant profile.
Traditionally, the selection of profiles to be monitored
has been carried out manually. Appropriate candidates are
selected, for example through research on the Internet or other
investigations. In this way, however, new or short-term profiles
are hardly detected. Here, automated approaches can help in
the long-term.
For example, the task of automatically identifying a danger-
ous militant profile, associated with a certain crime area, given
a group of profiles can be interpreted as a classification task.
Let Pbe the set of all profiles of a particular social network,
and Rthe set of risk classes, corresponding to an area of crime.
Then the selection of potential militants profiles is a surjective
mapping f:RP. An overview of classification techniques
(supervised learning methods) is given, for example, in [22]
[23]. However, as already emphasized by [24], a large amount
of training data is needed to train classifiers with sufficient
accuracy. This problem can be addressed, for example, by
the use of semi-supervised learning methods, such as self-
training or co-training. An overview of methods is described,
e. g., in [25]. Whichever method is chosen, the performance
depends on the choice of appropriate features. These should
generally have sufficient discrimination power and should be
as independent as possible.
Particularly in the context of social networks, the use of
techniques for recommender machines is often used (push-
mode) instead of classification (pull mode). Typically, such
systems use Collaborative Filtering [26] [27], Content-based
Filtering [28] [27], or a combination of both. In recent years,
a whole series of studies have been devoted to the creation of
friendships in social networks using these classic approaches
[29] [30]. More recent approaches are based on social graphs
[31] [32] or semantic analyzes, especially LDA, which attempt
to produce recommendations based on lifestyles [33] [34]
[35] [36]. However, the inclusion of structured data is more
reliable than the analysis of latent topics and is therefore
more suitable for classifying threats. Naruchitparames et al.
presented a recommender system based on genetic algorithms
[37]. As a feature (social genome), they propose the following
Facebook feature:
common friends,
location,
age range,
common interests (likes and music),
photo tags,
events,
groups,
movies,
education,
religious and political attitude.
Manca et al. criticize earlier approaches because they do
not take into account a mutual interest which is, however,
necessary for friendship. They suggest a similarity-based rec-
ommender as a basis for friendships using so-called Social
Bookmarks, i. e., shared bookmarks on the Internet [38]. Tags
of shared images are the basis for the recommender system
proposed by Cheung et al. and are another interesting feature to
generate recommendations in social networks [39]. In a similar
manner, a general classifier can be trained based on the profiles
of known dangerous militants or offenders. For example,
by means of corresponding known profiles, a classifier or a
recommendation system could be implemented for detecting
profiles of the hooligan scene or radical political groupings.
Adapting this approach, a classifier can be trained in the
sense of supervised learning, which can automatically detect
such dangerous militants profiles. We can use a social feature
vector ~
fsfor each profile (see Equation (1)) as a basis for the
computational task of the classification and recommendation
of dangerous militants.
~
fs=
friends
location
age
interests
· · ·
(1)
Considering this as a binary classification task, we need
to assign each profile ~
fseither to the class of potential dan-
gerous militants profiles or not. Assuming the features fs
iare
independent, we can use the Bayes theorem for computation
(see 2).
ˆy= arg max
ci∈{0,1}p(ci)
|~
fs|
Y
j=0
p(fs
j|ci)(2)
Although we know that this assumption is not true, experi-
ences have shown that this approach still produces good results
[40]. In general, supervised approaches need a sufficiently
large set of training examples which is a problem in many
cases. To overcome this, we can use a bootstrapping approach,
as shown in [41].
B. Assessment of the risk of dangers
After the potential dangerous militants profiles have been
selected, the content analysis of the communication takes
151
International Journal
o
n Advances in Intell
igent Systems, vol
10
no
3
&
4
1
7
, http://www.iariajournals.org/intelligent_sys
tems/
201
7
, © Copyright by authors, Published under agreement with IARIA
-
www.iaria.org
0 1000 3000 5000 7000
# negative comments
Jun 15 Jan 16
Dec 15
Jul 15 Aug 15 Sep 15 Oct 15 Nov 15
Figure 2: Results of a short-term study on the development of sentiment on the Facebook page of Pegida e.V. between June
2015 and January 2016. The blue areas mark the 95 %prediction interval. Red lines denote actual incidents during this period
of time. The gray area marks a period with missing data.
place. This step is necessary to determine whether the extrac-
tion of further information is necessary to elucidate various
modalities (location, time, participants, etc.) of possible events.
A prerequisite for the assessment of the probability that the
danger occurs is once again the experience-based knowledge
of the investigator, which must be available for each individual
risk type, for example, in the FoTM as discussed in [20] [41].
After defining the risk classes risk1, ..., risknwhich
should be monitored, the explicit definition of the corre-
sponding danger topics is made: Θrisk =ϑrisk1, ..., ϑriskn.
A risk class describes the amount of all offenses belong-
ing to a defined group, for example, left or right-winged
politically motivated crimes. A risk topic includes all the
terms and associations that characterize such a risk class.
Afterwards, the selection of potential or known dangerous
militants profiles leads to a set of candidate profiles for each
risk class Pc
i=pi1, ..., pik P, i = 1, ..., n from the set of
available profiles of the investigated social network is carried
out taking into account a particular risk class to limit the
scope of observation and analysis. Subsequently, the topics
Θcom =ϑcom1, ..., ϑcomnof the communication between these
profiles must be extracted and it must be then analyzed to what
extend they overlap with the risk topics. Afterwards, they are
evaluated. In the simplest case the overlap can be represented
binarily as shown in Equation (3).
f(Θcom) = 1if Θcom Θrisk 6=
0otherwise (3)
In order to quantify the degree of correspondence of Θcom
and Θrisk , a corresponding metric is needed to compare
probability distributions over the terms tof a topic. Niekler
and J¨
ahnichen examined the suitability of the Jensen-Shannon
divergence, the cosine similarity, and the dice coefficient as a
measurement of similarity for various topics [42]. As a result,
it was found that the best results were obtained on the basis of
the cosine similarity sim(ϑcom, ϑrisk ). Adapted to the present
application, sim(ϑcom, ϑrisk )is thus defined as:
sim(ϑcom, ϑrisk ) = ϑcom ·ϑrisk
kϑcomkkϑrisk k(4)
If f(Θcom) = 1, i. e., ϑcomi|ϑcomiΘrisk , the analysis
of the sentiment Sin the network is carried out. Approaches
are found in the literature, especially for Twitter messages [43]
[44]. In principle, these approaches can also be applied to other
social networks such as Facebook. If the sentiment exceeds a
threshold value ε, an increased risk can be assumed.
To evaluate this hypothesis the communication on the
Facebook page of “Pegida e. V.” (a mostly right-winged or-
ganization in Germany) was analyzed over a period of eight
months, between June 2015 and January 2016. The extracted
textual communication data was divided into individual sen-
tences (tokenization). Subsequently, one out of three polarity
classes pol: positive (+), negative ()or neutral (0) was
assigned to each sentence susing a probabilistic language
model. Equations (5) and (6) show the associated likelihood
function and the derived scoring function.
log2P(s, pol) = log2P(s|pol) + log2P(pol)(5)
score(s, pol) = log2P(s, pol)
|s|+ 2 (6)
The polarity class with the highest score is assigned to the
respective sentence. The “Multi-Domain Sentiment Encyclo-
pedia for the German Language”, which was developed at the
Darmstadt University of Applied Sciences, formed the basis for
the training. It contains extracted mood-bearing terms from the
MLSA-A corpus [45], the pressrelations dataset [46], and the
152
International Journal
o
n Advances in Intell
igent Systems, vol
10
no
3
&
4
1
7
, http://www.iariajournals.org/intelligent_sys
tems/
201
7
, © Copyright by authors, Published under agreement with IARIA
-
www.iaria.org
“German Sentiment Vocabulary” (SentiWS) [47], all annotated
with the average polarity values in the range [1,1].
The sentiment of a message m={s1, ..., sn}(post,
comment) is decided in the simplest case by the number
of its positive sentences s+and/or negative sentences s.
The sentiment that dominates the constituent sentences also
determines the sentiment of the whole message (see Equation
(7)). In case of equality, the message is considered to be neutral
m0.
S(m) =
m+if |s+|>|s|
mif |s+|<|s|
m0otherwise
(7)
This approach, of course, is only a rough estimate of the
sentiment, since it does not take into account the connection
between meaning (semantic) and sentiment of a sentence.
The accuracy, however, appears sufficient for a first check of
the hypothesis, since the messages themselves were filtered
in advance by the topic analysis. The results are presented
in a histogram (see Fig. 2), with only negative messages
(comments) mbeing taken into account.
Comparing the development of the sentiment of the com-
ments in the network with the events during this period
(marked by red lines), it was found that there is a possible
correlation between these two. For example, on January 11th,
2016 serious riots lead by the right-winged scene occurred
during the demonstration of the sister organization Legida
e. kV. in Leipzig (Germany). Members in the Pegida network
were also encouraged to attend this event. Similar to most
of the cases, it can be clearly seen that the peak of negative
communication is situated immediately before the incident.
The sudden reduction in conversations at the time of the event
can be explained by the active participation of the members in
the event. The 95%-prediction interval (blue lines) supports the
assumption that incidents mostly occur after a local or global
peak.
Even if this short study is not considered representative and
a random correlation between the occurrence of the incident
and the discussion in the network cannot be ruled out, it still
shows the potential of the presented approach. At this point,
additional long-term studies with larger data sets considering
different networks are necessary.
C. Detection of Leaders and Multipliers
Leaders and multiplier in the context of the intended anal-
ysis of social networks are individuals, who exert a significant
amount of influence on the opinion and sentiment of other
users of the network through their actions. In social sciences
the term ’opinion leader’ was introduced before 1957 by
Katz and Lazarsfeld’s research on diffusion theory [48]. Their
proposed two-step flow model (see Fig. 3) retains validity in
the digital age, especially in the context of social media.
Katz et al. assume that information disseminated in the
Social Network is received, strengthened and enriched by
opinion-leaders Liin their social environment. Since opinion-
leadership is strongly knowledge-driven and thus topic-
dependent, this model must be supplemented by various the-
matically limited opinion-leaders Lϑi. Each individual is then
influenced by a variety of heterogeneous opinion leaders in his
opinion as illustrated in Fig. 3. This signifies, that the opinion
of an individual is mostly formed by its social environment. In
1962, Rogers references these ideas and defines opinion leader
as follows:
“Opinion leadership is the degree to which an
individual is able to influence informally other indi-
viduals’ attitudes or overt behavior in a desired way
with relative frequency.” [49, p. 331]
Social Media
𝐿𝜗1 𝐿𝜗2 𝐿𝜗3 𝐿𝜗4
Figure 3: Extended two-step flow model adapted from [48]. In-
formation is spread throughout social media. Individuals with
a high level of competence at strategic local positions receive
and amplify this information according to their competence
(opinion-leader Lϑ1, ..., Lϑ4) and spread it to its followers and
friends. This means, each individual’s opinion is influenced
by different opinion leaders depending on the topic (different
colors) discussed in the network.
For the present study, the most important question to
answer is what influence means, or rather how to identify an
opinion leader or how the influencer can be distinguished from
those being influenced. Katz defined the following features
[48]:
1) personification of certain values
2) competence
3) strategic social location
One approach to identify opinion-leaders is to extract and
analyze the content of nodes and edges of networks to mine
leadership features. For instance, the sentiment of communi-
cation pieces can be analyzed to detect the influence of their
authors, as shown by Huang et. al., who aim to detect the
most influential comments in a network this way [50]. Another
strategy is to perform topic mining to categorize content and
detect opinion-leaders for each topic individually, as opinion-
leadership is context-dependent [48] [51]. For this purpose,
Latent Dirichlet Allocation (LDA) [52] can be used, as seen
in the work of [53]. We considered the implementation of
content-based methods problematic, as texts in social networks
lack correct spelling and formal structure, which impairs such
methods’ performance.
Another approach to identifying leaders is to analyze the
flow of information in a network. By monitoring how the in-
teraction of actors evolves over time, one can identify patterns
and individuals of significance within them. To achieve this,
some model of information propagation is required, such as the
153
International Journal
o
n Advances in Intell
igent Systems, vol
10
no
3
&
4
1
7
, http://www.iariajournals.org/intelligent_sys
tems/
201
7
, © Copyright by authors, Published under agreement with IARIA
-
www.iaria.org
Markov processes used by [54] and the probabilistic models
proposed by [55]. These interaction-based methods consider
both topological features and their dynamics over time. How-
ever, the latter are not yet considered by our framework and
are reserved for future developments.
We utilized methods, which are solely based on a network’s
topology, therefore, consider features, such as node degree,
neighborhood distances and clusters, to identify opinion lead-
ers. One implementation of this is the calculation of node
centrality. The underlying assumption is, that the more influ-
ence an individual gains, the more central it is in its network.
Which centrality measure is most suitable is dependent on the
application domain. We judged eigenvector centrality to be
most adequate, specifically Google’s PageRank algorithm [56],
which functions in a similar fashion. It recursively assigns a
rank Rto each node A, based on the rank of the nodes of its
incoming edges Tiand its total number of links Ci. The value
of an edge is strongly dependent on the score of its originator
(see Equation (8)).
R(A) = 1d
N+d
n
X
i=1
R(T1)
C(T1),0d1(8)
With the damping factor d, normalized over the number
of all nodes of the network N, a part of the resulting rank
can be subtracted and distributed to all nodes. The application
of PageRank for the purposes of opinion leader detection has
seen merely moderate success [57] [58]. With LeaderRank, L¨
u
et al. advocate further development and optimization of this
algorithm for social networks, and have achieved surprisingly
good results [59]. Users are considered as nodes and directed
edges as relationships between opinion leaders and users. All
users are also bidirectionally connected to a ground node. At
time step t0, all nodes receive the score si(0) = 1 except
for the ground node initialized with sg(0) = 0. Equation (9)
describes the process of probability flow through the network,
where si(t)indicates the LeaderRank score of a node iat time
step t.
si(t+ 1) =
N+1
X
j=1
aji
kout
j
sj(t)(9)
Depending on whether or not there exists a directed edge
from node ito node j, the value 0or 1is assigned to aij.
kout
idescribes the number of outgoing edges of a node. The
final score is obtained as the score of the respective node at
the convergence time tcand the base node score at the same
time, as shown in Equation (10).
Si=si(tc) + sg(tc)
N(10)
The advantage of this algorithm compared to PageRank
is that the convergence is faster and above all that nodes,
that spread information faster and further, can be found. In
later work, for example, by introducing a weighting factor, as
in [60] or [61], susceptibility to noisy data has been further
reduced and the ability to find influential distributors (hubs) of
information has been added.
However, there might be cases in which LeaderRank would
assign high scores to individuals, which are not relevant for
the present application. When a user attained a significant
audience, while also actively following many opinion leaders,
we argue that their influence is based on their activity in
the network and not their opinion, as their presence makes
them more likely to be followed. We propose an approach
to eliminate such hybrid leaders from the top ranks, which
punishes the LeaderRank score LR(ϑi)of users with many
interactions in the network, meaning, those users who follow
many leaders. This way the top ranked users are pure leaders,
whose influence is purely based on their opinions.
P SC (Lϑi) = LR(Lϑi)
1 + kout
i
kout
total
LRtotal
(11)
One way to calculate the PureScore of a particular topic-
specific opinion leader P SC (Lϑi)is shown in Equation (11).
The PureScore of a certain topic-specific opinion-leader is
calculated by dividing its original LeaderRank score LR(Lϑi)
by a percentage of the maximum score (equal to the number
of users) defined by the node’s share of network activity, kout
being the number of outgoing links. However, this approach
needs to be evaluated in later work.
D. Visualization
If, with the approach described above, a risk greater than a
threshold value εwas determined, further information such
as locations, times and people involved are extracted from
the text and subsequently transferred to a corresponding map
with the help of geographical coordinates. An additional score
frisk (ϑ, Sϑ,|P|ϑ)provides information about the extent of the
expected risk, estimated from the risk class, the sentiment score
associated with it and the number of people involved in that
particular discussion. This value can, for example, be used to
color the geo location on a map, corresponding to a heat-scale.
The obtained result directly supports the short-term strategic
planning of security forces as proposed at the very beginning
of this section.
V. PROTOTYPICAL IMPLEMENTATION
The aim of the prototype’s architecture is to implement the
frameworks described in [20] as well as the sections above. It
was programmed with Java and built as an Eclipse Rich Client
Platform (RCP). Its OSGi implementation Equinox allows for
a service-oriented architecture, consisting of three tiers:
1) Persistence: Data is fetched from the various social-
network databases and put into EMF models. The
models are stored into a, as for now, local EMF Store
server. Thus, the databases and the EMF Store server
form this tier. Any annotations and meta-data are also
held by the models.
2) Logic-Tier: This tier contains various linguistic ser-
vices, e. g., for topic modeling, used for annotation
and querying. The modeling service, which provides
CRUD-operations (Create, Read, Update, Delete) for
models in the EMF Store server, also resides here.
154
International Journal
o
n Advances in Intell
igent Systems, vol
10
no
3
&
4
1
7
, http://www.iariajournals.org/intelligent_sys
tems/
201
7
, © Copyright by authors, Published under agreement with IARIA
-
www.iaria.org
Furthermore, all data retrieval services, which com-
municate with corresponding social-network APIs,
are part of this tier.
3) Access: The high-level services, usually directly con-
trollable by the UI, define this tier. At the current state
of the development, this is the retrieval service, which
initializes data fetching from the social networks, the
query-service, used for data retrieval from the EMF
Store server and the annotation services, which use
the linguistic services to enrich models.
Logic PersistenceAccess
Retrieval Service
EMF Store Server
Query Service
Model Service
Annotation Services
Linguistic Services
Social Network
Databases
Facebook
Graph Database ...
Social Network
Retrieval services
Facebook
Retrieval Service
Figure 4: SoNA architecture overview with the respective
services.
Figure 4 provides a visual representation of these tiers.
When developing this application further, efforts will be made
to make the prototype more closely resemble the SoNA
framework described in this work. Permission services can be
realized through user profiles in the EMF Store server.
Figure 5: The user interface of the prototype. Shown is a down-
loaded Facebook page visualized in a graph. The Facebook
logo node represents the page, which is connected to posts,
which in turn are connected to their comments. The outermost
nodes are users, which are associated to the comments they
created.
The prototype provides a user interface to retrieve data
from Social Networks, currently Facebook, and visualizes it
in a 2D graph, as shown in Figure 5. Several filter options
are available in order to reduce the output network to the most
relevant nodes depending on the investigator’s needs. Data can
be retrieved from a specific part of the social network for a
certain period of time or a certain amount of content (i. e.,
posts and comments), before being stored in models. Created
models can then be annotated following the process chain as
discussed in the former sections.
VI. CONCLUSION AND FUTURE WORK
In this paper the theoretical framework SoNA was pre-
sented, which allows investigators of law enforcement agencies
as well as intelligence services to monitor social networks
in order to gather information about potentially dangerous
activities. This information can support the long- and short-
term planning of the deployment of security forces. It was
shown that the knowledge gained by applying this framework
can directly increase the resilience of a society in the first
two stages of the resilience cycle. Furthermore, given the
complexity of the language used in Social Media it was nec-
essary to apply a knowledge-based and word-based approach.
In this respect, a process chain for analyzing social networks
was proposed and the main steps were discussed in detail.
These include the selection of dangerous militant profiles,
the assessment of the risk and the detection of leaders and
multipliers. We presented a prototype of the SoNA framework,
which aims to implement these aspects. Future work will
include the remaining steps of the process chain as well as
the evaluation of the entire framework. In order to do so, it
is fundamental to create an appropriate test environment in
collaboration with law enforcement agencies.
REFERENCES
[1] M. Spranger, F. Heinke, S. Grunert, and D. Labudde, “Towards pre-
dictive policing: Knowledge-based monitoring of social networks,
in Proceedings of the Fifth International Conference on Advances
in Information Mining and Management (IMMM 2015), IARIA, Ed.
ThinkMind Library, 2015.
[2] ZIS, “Jahresbericht 2013/14,” 2014, [Online]
https://lzpd.polizei.nrw/sites/default/files/2016-12/13-
14 Jahresbericht.pdf, last accessed 2017-11-29.
[3] Deutsches Institut f¨
ur Normung, “Begriffe im Rettungswesen,” April
2015.
[4] J. Walker and M. Cooper, “Genealogies of resilience from systems ecol-
ogy to the political economy of crisis adaptation.” Security Dialogue,
vol. 42, no. 2, 2011, pp. 143–160.
[5] C. Edwards, Resilient Nation. London: Demos, 2009.
[6] K. Thoma, Ed., Resilien-Tech: Resilience-by-Design; Strategie f¨
ur die
technologischen Zukunftsthemen, ser. acatech STUDIE. M¨
unchen:
Utz, 2014.
[7] S. L. Pimm, The balance of nature? Ecological issues in the conser-
vation of species and communities, 2nd ed. Chicago u.a.: Univ. of
Chicago Press, 1992.
[8] C. S. Holling, “Engineering resilience versus ecological resilience,”
Engineering within ecological constraints, 1996, pp. 31–44.
[9] B. Pearsall, “Predictive policing: The future of law enforcement?” NIJ
Journal, no. 266, 2009.
[10] K. Thoma, “Resilien-Tech. Resilience-by-Design: Strategie f¨
ur die
technologischen Zukunftsthemen,” acatech Studie, 2014.
[11] C. D¨
urscheid, F. Wagner, and S. Brommer, Wie Jugendliche schreiben:
Schreibkompetenz und neue Medien, ser. Linguistik - Impulse &
Tendenzen. Berlin u.a.: de Gruyter, 2010, vol. 41.
155
International Journal
o
n Advances in Intell
igent Systems, vol
10
no
3
&
4
1
7
, http://www.iariajournals.org/intelligent_sys
tems/
201
7
, © Copyright by authors, Published under agreement with IARIA
-
www.iaria.org
[12] M. Zappavigna, Discourse of Twitter and social media: [how we use
language to create affiliation on the web], ser. Continuum discourse
series. London u.a.: Continuum Publ, 2012.
[13] C. D¨
urscheid, “Medienkommunikation im Kontinuum von M¨
undlichkeit
und Schriftlichkeit. Theoretische und empirische Probleme,” Zeitschrift
f¨
ur Angewandte Linguistik, vol. 38, 2003, pp. 37–56.
[14] S. Hintze, Emotionalit¨
atsmarker in Kommentaren auf der PEGIDA-
Facebook-Seite, ser. Networx, 2015, vol. 71.
[15] M. Beißwenger, Kommunikation in virtuellen Welten: Sprache, Text und
Wirklichkeit ; eine Untersuchung zur Konzeptionalit¨
at von Kommunika-
tionsvollz¨
ugen und zur textuellen Konstruktion von Welt in synchroner
Internet-Kommunikation, exemplifiziert am Beispiel eines Webchats.
Stuttgart: Ibidem-Verl., 2000.
[16] K. Luckhardt, “Stilanalysen zur Chat-Kommunikation: Eine korpus-
gest¨
utzte Untersuchung am Beispiel eines medialen Chats,” Ph.D.
dissertation, TU Dortmund, 2009.
[17] M. O’Donnell, “The UAM CorpusTool: Software for corpus annotation
and exploration,” in Bretones Callejas (Hg.) 2009 – Applied linguistics
now, pp. 1433–1447.
[18] C.-V. Schnitzer, “Linguistische Aspekte der Kommunikation in
den neueren elektronischen Medien SMS-E-Mail-Facebook,” Doktor,
Ludwig-Maximilians-Universit¨
at, M¨
unchen, 2012.
[19] E. Dresner and S. C. Herring, “Functions of the Nonverbal in CMC:
Emoticons and Illocutionary Force,” Communication Theory, vol. 20,
no. 3, 2010, pp. 249–268.
[20] M. Spranger, S. Schildbach, F. Heinke, S. Grunert, and D. Labudde,
“Semantic tools for forensics: A highly adaptable framework,” in Proc.
2nd. International Conference on Advances in Information Management
and Mining (IMMM). ThinkMind Library, 2012, pp. 27–31.
[21] B.-D. 16/3570, “Schriftliche Fragen mit den in der Woche vom
20. November 2006 eingegangenen Antworten der Bundesregierung,
Drucksache des Deutschen Bundestages 16/3570 vom 24. November
2006, 2006.
[22] S. B. Kotsiantis, “Supervised machine learning: A review of classifica-
tion techniques,” in Proceedings of the 2007 Conference on Emerging
Artificial Intelligence Applications in Computer Engineering: Real
Word AI Systems with Applications in eHealth, HCI, Information
Retrieval and Pervasive Technologies. Amsterdam, The Netherlands,
The Netherlands: IOS Press, 2007, pp. 3–24.
[23] J. Han, M. Kamber, and J. Pei, Data Mining: Concepts and techniques,
3rd ed., ser. The Morgan Kaufmann series in data management systems.
Waltham Mass.: Morgan Kaufmann, 2012.
[24] M. Ikonomakis, S. Kotsiantis, and V. Tampakas, “Text classification us-
ing machine learning techniques,” WSEAS Transactions on Computers,
vol. 4, no. 8, 2005, pp. 966–974.
[25] O. Chapelle, B. Sch¨
olkopf, and A. Zien, Semi-supervised learning, ser.
Adaptive computation and machine learning, 2006.
[26] D. Goldberg, D. Nichols, B. M. Oki, and D. Terry, “Using collaborative
filtering to weave an information tapestry,” Communications of the
ACM, vol. 35, no. 12, 1992, pp. 61–70.
[27] F. Ricci, L. Rokach, and B. Shapira, “Introduction to Recommender
Systems Handbook,” in Recommender Systems Handbook, F. Ricci,
L. Rokach, B. Shapira, and P. B. Kantor, Eds. New York: Springer,
2011, pp. 1–35.
[28] R. van Meteren and M. van Someren, “Using content-based filtering for
recommendation,” in Proceedings of the Machine Learning in the New
Information Age: MLnet/ECML2000 Workshop, 2000, pp. 47–56.
[29] L. Bian and H. Holtzman, “Online friend recommendation through
personality matching and collaborative filtering,” Proc. of UBICOMM,
2011, pp. 230–235.
[30] V. Agarwal and K. K. Bharadwaj, “A collaborative filtering framework
for friends recommendation in social networks based on interaction
intensity and adaptive user similarity,” Social Network Analysis and
Mining, vol. 3, no. 3, 2013, pp. 359–379.
[31] N. B. Silva, R. Tsang, G. D. C. Cavalcanti, and J. Tsang, “A graph-
based friend recommendation system using genetic algorithm,” in
Evolutionary Computation (CEC), 2010 IEEE Congress on, 2010, pp.
1–7.
[32] F. Akbari, A. H. Tajfar, and A. F. Nejad, “Graph-based friend rec-
ommendation in social networks using artificial bee colony,” in De-
pendable, Autonomic and Secure Computing (DASC), 2013 IEEE 11th
International Conference on, 2013, pp. 464–468.
[33] N. M. Eklaspur and A. S. Pashupatimath, “A friend recommender
system for social networks by life style extraction using probabilistic
method-friendtome,” International Journal of Computer Science Trends
and Technology (IJCST), vol. 3, no. 3, 2015.
[34] Z. Wang, J. Liao, Q. Cao, H. Qi, and Z. Wang, “Friendbook: A
semantic-based friend recommendation system for social networks,”
IEEE Transactions on Mobile Computing, vol. 14, no. 3, 2015, pp.
538–551.
[35] T. R. Kacchi and A. V. Deorankar, “Friend recommendation system
based on lifestyles of users,” in Advances in Electrical, Electronics,
Information, Communication and Bio-Informatics (AEEICB), 2016 2nd
International Conference on, 2016, pp. 682–685.
[36] D. M. Jadhavar and V. R. Chirchi, “Friend recommendation system
for online social networks,” International Journal of Computer Appli-
cations, vol. 153, no. 12, 2016.
[37] J. Naruchitparames, M. H. G¨
une, and S. J. Louis, “Friend recommenda-
tions in social networks using genetic algorithms and network topology,
in 2011 IEEE Congress of Evolutionary Computation (CEC), 2011, pp.
2207–2214.
[38] M. Manca, L. Boratto, and S. Carta, “Producing friend recommenda-
tions in a social bookmarking system by mining users content,” in The
Third International Conference on Advances in Information Mining and
Management (IMMM 2013), IARIA, Ed. IARIA, 2013, pp. 59–64.
[39] M. Cheung and J. She, “Bag-of-features tagging approach for a better
recommendation with social big data,” in Proc. 4th. International
Conference on Advances in Information Mining and Management.
ThinkMind Library, 2014, p. 83 to 88.
[40] I. Rish, “An empirical study of the Na¨
ıve Bayes classifier,” IJCAI 2001
Workshop on Empirical Methods in Artificial Intelligence, vol. 22, no. 3,
pp. 41–46.
[41] M. Spranger and D. Labudde, “Semantic tools for forensics: Approaches
in forensic text analysis,” in The Third International Conference on
Advances in Information Mining and Management (IMMM 2013),
IARIA, Ed. IARIA, 2013, pp. 97–100.
[42] A. Niekler and P. J¨
ahnichen, “Matching Results of Latent Dirichlet Al-
location for Text,” in Proceedings of the 11th International Conference
on Cognitive Modeling, ICCM 2012, 2012.
[43] X. Wan, “Co-training for cross-lingual sentiment classification,” in
Proceedings of the Joint Conference of the 47th Annual Meeting of the
ACL and the 4th International Joint Conference on Natural Language
Processing of the AFNLP: Volume 1-volume 1, 2009, pp. 235–243.
[44] S. M. Mohammad, S. Kiritchenko, and X. Zhu, “NRC-Canada: Building
the state-of-the-art in sentiment analysis of tweets,” arXiv preprint
arXiv:1308.6242, 2013.
[45] S. Clematide, S. Gindl, M. Klenner, S. Petrakis, R. Remus, J. Rup-
penhofer, U. Waltinger, and M. Wiegand, “MLSA-A Multi-layered
Reference Corpus for German Sentiment Analysis,” in LREC, 2012,
pp. 3551–3556.
[46] T. Scholz, S. Conrad, and L. Hillekamps, “Opinion mining on a german
corpus of a media response analysis,” in International Conference on
Text, Speech and Dialogue, 2012, pp. 39–46.
[47] R. Remus, U. Quasthoff, and G. Heyer, “SentiWS-A Publicly Available
German-language Resource for Sentiment Analysis,” in LREC, 2010.
[48] E. Katz, “The two-step flow of communication: An up-to-date report on
an hypothesis,” Public Opinion Quarterly, vol. 21, no. 1, Anniversary
Issue Devoted to Twenty Years of Public Opinion Research, 1957, p. 61.
[49] E. M. Rogers, Diffusion of innovations. New York: The Free Press,
1962.
[50] B. Huang, G. Yu, and H. R. Karimi, “The finding and dynamic
detection of opinion leaders in social network,” Mathematical Problems
in Engineering, vol. 2014, 2014, pp. 1–7.
[51] P. Parau, C. Lemnaru, M. Dinsoreanu, and R. Potolea, “Opinion leader
detection.” in Sentiment analysis in social networks, F. A. Pozzi,
E. Fersini, E. Messina, and B. Liu, Eds., 2016, pp. 157–170.
[52] D. M. Blei, A. Y. Ng, and M. I. Jordan, “Latent Dirichlet Allocation,”
J. Mach. Learn. Res., vol. 3, 2003, pp. 993–1022.
156
International Journal
o
n Advances in Intell
igent Systems, vol
10
no
3
&
4
1
7
, http://www.iariajournals.org/intelligent_sys
tems/
201
7
, © Copyright by authors, Published under agreement with IARIA
-
www.iaria.org
[53] X. Song, Y. Chi, K. Hino, and B. Tseng, “Identifying opinion leaders in
the blogosphere,” in Proceedings of the sixteenth ACM conference on
Conference on information and knowledge management - CIKM ’07,
M. J. Silva, A. O. Falc˜
ao, A. A. F. Laender, R. Baeza-Yates, D. L.
McGuinness, B. Olstad, and Ø. H. Olsen, Eds. New York, New York,
USA: ACM Press, 2007, p. 971.
[54] B. Amor, S. Vuik, R. Callahan, A. Darzi, S. N. Yaliraki, and M. Bara-
hona, “Community detection and role identification in directed net-
works: Understanding the Twitter network of the care.data debate,”
CoRR, vol. abs/1508.03165, 2015.
[55] M. Richardson and P. Domingos, “Mining knowledge-sharing sites
for viral marketing,” in Proceedings of the eighth ACM SIGKDD
international conference on Knowledge discovery and data mining,
Za&iuml and O. R. ane, Eds. New York, NY: ACM, 2002, p. 61.
[56] S. Brin and L. Page, “The anatomy of a large-scale hypertextual web
search engine,” Comput. Netw. ISDN Syst., vol. 30, no. 1-7, Apr. 1998,
pp. 107–117.
[57] C. Egger, “Identifying Key Opinion Leaders in Social Networks: An
Approach to use Instagram Data to Rate and Identify Key Opinion
Leader for a Specific Business Field,” Master Thesis, TH K¨
oln -
University of Applied Sciences, K¨
oln, 2016.
[58] M. Z. Shafiq, M. U. Ilyas, A. X. Liu, and H. Radha, “Identifying leaders
and followers in online social networks,” IEEE Journal on Selected
Areas in Communications, vol. 31, no. 9, 2013, pp. 618–628.
[59] L. L¨
u, Y.-C. Zhang, C. H. Yeung, and T. Zhou, “Leaders in social
networks, the Delicious case,” PLoS One, vol. 6, no. 6, 2011, p. e21202.
[60] Q. Li, T. Zhou, L. L¨
u, and D. Chen, “Identifying influential spreaders
by weighted LeaderRank,” Physica A: Statistical Mechanics and its
Applications, vol. 404, no. Supplement C, 2014, pp. 47–55.
[61] Z. H. Zhang, G. P. Jiang, Y. R. Song, L. L. Xia, and Q. Chen,
“An improved weighted leaderrank algorithm for identifying influential
spreaders in complex networks,” in 2017 IEEE International Conference
on Computational Science and Engineering (CSE) and IEEE Interna-
tional Conference on Embedded and Ubiquitous Computing (EUC),
vol. 1, 2017, pp. 748–751.
... k out i describes the number of outgoing edges of a node. The final score is obtained as the score of the respective node at the convergence step t c and the obtained ground node score, as shown in (2). At t c , equilibration of LeaderRank scores towards a steady state can be observed. ...
... [2308] [2] [2313] ...
Conference Paper
Full-text available
In recent years, the automated, efficient and sensitive monitoring of social networks has become increasingly important for criminal investigations and crime prevention. Previously, we have shown that the detection of opinion leaders is of great interest in forensic applications. In the present study, it is argued that state of the art opinion leader detection methods have weaknesses if networks exhibit star-like social graph topology, whereas these topologies result from the interaction of users with similar interests. This is typically the case for Facebook pages of political organizations. In these cases, the underlying topologies are highly focused on one (or only a few) central actor(s) and lead to less meaningful results by classic measures of node centrality commonly used for leader detection. The presents study examines these aspects closer and exemplifies them with the help of data collected from the Facebook page of a German political party for five consecutive months. Furthermore, a quantitative indicator for describing star-like network topologies is introduced and discussed. This measure can be of great value in assessing the applicability of established leader detection methods. Finally, a modified LeaderRank score is proposed -- the CompetenceRank -- which aims to address discussed problems.
... The final score is obtained as the score of the respective vertex at the convergence step t c and the obtained ground vertex score, as shown in (4). At t c , equilibration of LeaderRank scores towards a steady state is observed. ...
Article
Full-text available
In recent years, the automated, efficient and sensitive monitoring of social networks has become increasingly important for the criminal investigation process and crime prevention. Previously, we have shown that the detection of opinion leaders is of great interest in forensic applications to gather important information. In the current work, it is argued that state of the art methods, determining the relative degree to which an opinion leader exerts influence over the network, have weaknesses if networks exhibit a star-like social graph topology, whereas these topologies result from the interaction of users with similar interests. This is typically the case in networks of political organizations. In these cases, the underlying topologies are highly focused on one (or only a few) central actor(s) and lead to less meaningful results by classic measures of node centrality commonly used to ascertain the degree of leadership. With the help of data collected from the Facebook and Twitter network of a German political party, these aspects are examined and a quantitative indicator for describing star-like network topologies is introduced and discussed. This measure can be of great value in assessing the applicability of established leader detection methods. Finally, two variations of a new measure– the CompetenceRank – which is based on the LeaderRAnk score and aims to address the discussed problems in cases with and without additional network data such as likes and shares, are proposed.
... Text Mining, also referred to as Text Analysis, focuses on the analysis of texts in order to receive high level information and latent patterns. It can be applied to many different areas [3], however, it plays a special role in forensics and predictive policing, where it can be used to detect events with a potential to escalate [4]. Event detection is a specific Text Mining problem in which texts are analyzed in order to mine a set of texts that have a semantic link or share conceptual patterns regarding past or future events [5] [6]. ...
Article
Full-text available
Announcements of events are regularly spread using the Internet, e.g., via online newspapers or social media. Often, these events involve playing music publicly that is protected by international copyright laws. Authorities entrusted with the protection of the artists' interests have to find unregistered music events in order to fully exercise their duty. As a requirement, they need to find texts in the Internet that are related to such events like announcements or reports. However, event detection is a challenging task in the field of Text Mining due to the enormous variety of information that needs to be considered and the large amount of data that needs to be processed. In this paper, a process chain for the detection of music events incorporating external knowledge is proposed. Furthermore, a feature selection algorithm based on ant colony optimization to find featurse with a high degree of explanatory power is presented. Finally, the performance of five different machine learning algorithms including two learning ensembles is compared using various feature sets and two different datasets. The best performances reach an F1-measure of 0.95 for music texts and 0.968 for music event texts, respectively.
Chapter
Full-text available
With the rise of social media as an important channel for the debate and discussion of public affairs, online social networks such as Twitter have become important platforms for public information and engagement by policy makers. To communicate effectively through Twitter, policy makers need to understand how influence and interest propagate within its network of users. In this chapter we use graph-theoretic methods to analyse the Twitter debate surrounding NHS England's controversial care.data scheme. Directionality is a crucial feature of the Twitter social graph - information flows from the followed to the followers - but is often ignored in social network analyses; our methods are based on the behaviour of dynamic processes on the network and can be applied naturally to directed networks. We uncover robust communities of users and show that these communities reflect how information flows through the Twitter network. We are also able to classify users by their differing roles in directing the flow of information through the network. Our methods and results will be useful to policy makers who would like to use Twitter effectively as a communication medium.
Conference Paper
Full-text available
Increasing the resilience of the society against disorders, such as disasters, attacks or threatening groups, is a major challenge. Recent events highlight the importance of a resilient society and steps which are required to be taken in resilience engineering. textit{A priori} the optimal way to handle such adverse events is to prevent them, or at least provide appropriate courses of preparation. The essential requirement for every kind of preparation is information about relevant upcoming events. Such information can be gained for example from social networks and can form the basis for a long-term and short-term strategic planning by security forces. For that purpose, we here propose an application framework for knowledge-based social network monitoring, which aims at predicting short-term activities, as well as the long-term development of potentially dangerous groups. In this work, a theoretical outline of this approach is given and discussed.
Book
Full-text available
Today, young people write in their leisure time far more than they did 15 years ago. Most often they use the new media to do their writing. This book explores whether the frequent writing of short messages and e-mails and participation in chats and social networks like Facebook have an influence on writing in school. Are there any similarities and relationships between the texts written in school and the private texts? For the first time, based on comprehensive data from Swiss students, this book provides empirical answers to these questions.
Conference Paper
Full-text available
The interests of users are always important for person-alized content recommendations on friendships, events and media content from the social big data. However, those interests may not be specified, which makes the recommendations challenging. One of the possible solutions is to analyze the user's interests from the shared content, especially images with manually annotated tags. They are shared on online social networks such as Flickr and Instagram. However, the accuracy of the recommendation is greatly affected by the accuracy of the tag, which is not always reliable. This paper demonstrates how a bag-of-features (BoF)-based tagging approach can help to improve the accuracy of recommendations using an unsupervised algorithm. A set of auxiliary tags is used to represent user interests and, hence, the recommendation. The approach is evaluated with over 500 user and 200k images from Flickr. It is proven that by BoF tagging (BoFT), friendship recommendation is possible without friendship/tag information and the recall and the precision rate are improved by about 50% over using user tags.
Chapter
Identifying opinion leaders and assessing their influence is highly relevant in several domains. This chapter presents a comprehensive analysis of state-of-the-art opinion leader detection strategies in social networks and the associated challenges. We classify the approaches into four categories, depending on the types of available and observable data: network topology, interaction, content mining, and content and interaction. For each category, we present relevant approaches, the metrics and algorithms employed, and evaluation remarks. A critical discussion of the advantages, limitations, and suitability of the approaches is also provided. The potential for generating added value by detection of opinion leaders is illustrated with different application scenarios.
Conference Paper
Due to the increasing exponential growth in the use of social networking services, different social networking services have provided us new radical ways of making friends. Different recommendation system exist which are classified into two top level categories as object recommendation and link recommendation. In this paper, our focus will be on later one. The proposed system will work like a client-server application where the user which is requesting the query act as a client. Life documents of each user are collected from the client with the help of browser. In this phase, these collected data will be stored into a file either in semi-structured or structured format accordingly. The life styles of users are extracted by using either hadoop technology or SQL depending on the type of file as input to it. Then the concept of reverse indexing is used for easy retrieval of the desired data. Then with the help of graph data structure we can represent the relationship between users. As recommendation is based on different priorities like similar interest, similar blood group, nearby location, ranking is also one of the factor. So, the ranks of users are calculated using the pseudocode mentioned in this paper. Finally, client/user sends a query and server will respond a list of friends to the user/client (browser) accordingly.