DISCOVERING CRITICAL SITUATIONS IN ONLINE SOCIAL

NETWORKS

A Neuro Fuzzy Approach to Alert Marketing Managers

Carolin Kaiser, Sabine Schlick and Freimut Bodendorf

Department of Information Systems, University of Erlangen-Nuremberg, Lange Gasse 20, Nuremberg, Germany

Keywords: Opinion mining, Social network analysis, Neuro fuzzy model, Internet.

Abstract: More and more people are exchanging their opinions in online social networks and influencing each other. It

is crucial for companies to observe opinion formation concerning their products. Thus, risks can be recog-

nized early on and counteractive measures can be initiated by marketing managers. A neuro fuzzy approach

is presented which allows the detection of critical situations in the process of opinion formation and the

alerting of marketing managers. Rules for identifying critical situations are learned on the basis of the opi-

nions of the network members, the influence of the opinion leaders and the structure of the network. The

opinions and characteristics of the network are identified by text mining and social network analysis. The

approach is illustrated by an exemplary application.

1 INTRODUCTION

The number of people who are engaged in online

social networks is increasing steadily. Within these

networks, people are passing on information and

evaluations of products. By discussing with each

other they influence one another’s opinions and

purchasing behavior. It is important for companies

to monitor the development of online opinions

continously in order to detect risks at an early stage

and to take preventive actions. Thus, the spread of

negative opinions can be stopped and the compay’s

image can be saved from damage.

According to diffusion theory (Rogers 2003) not

only the characteristics of a product but also the

social network have a great impact on the spread of

opinions. Opinion leaders are in a position to

influence many members of the network. The

structure of the network, i.e. the relationships among

the network members, determines how fast opinions

disseminate.

Critical situations within social networks arise

when negative opinions are on the verge of being

spread and causing damage to the company’s image

or sales volume. The detection of critical situations

is very difficult since many factors must be consi-

dered. The opinions of the network members and the

power of the opinion leaders as well as the structure

of the network influence the future opinion devel-

opment. All of these factors must be taken into ac-

count in order to judge whether a situation is critical

or non-critical. Considerable experience in the fields

of online social networks and marketing is vital for

evaluating situations correctly. The automation of

this complex and also time-consuming task poses an

immense challenge to research. A warning system

should not only be able to assess situations correctly

but should also be easy for marketing managers to

understand so that they can apply it intuitively.

An neuro fuzzy approach is introduced which

detects critical situations in the process of opinion

formation by taking the overall social network into

account. The opinions of the networks members are

first recognized by methods coming from text

mining. The opinion leaders and the network

structure are then characterized by key figures

coming from social network analysis. Based on this

information, a fuzzy perceptron learns rules which

enable the discovery of critical situations and the

warning of marketing managers. These rules can be

easily interpreted by marketing managers.

2 RELATED WORK

Opinion mining on the Internet has recently become

a popular field of research. There are many papers

Kaiser C., Schlick S. and Bodendorf F..

DISCOVERING CRITICAL SITUATIONS IN ONLINE SOCIAL NETWORKS - A Neuro Fuzzy Approach to Alert Marketing Managers.

DOI: 10.5220/0003070900560064

In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2010), pages 56-64

ISBN: 978-989-8425-28-7

 2010 SCITEPRESS (Science and Technology Publications, Lda.)

which apply text mining to online discussions in

order to reveal consumer opinions about products or

product features (Kaiser 2009, Dave, Lawrence and

Pennoc 2003, Pang, Lee and Vaithyanathan 2002,

Popescu and Etzioni 2005). Glance et. al (2005)

integrate several mining methods to enable online

opinion tracking. Kim und Hovy (2007) propose a

similar approach for predicting election results by

analyzing predictive opinions. All of these ap-

proaches only take a static view of online opinions.

Several papers focus on the dynamic evolution of

online activities. Viermetz, Skubacz, Ziegler and

Seipel (2008) monitor the evolution of short term

topics and long term trends. The system of Tong and

Yager (2004) automatically summarizes the devel-

opment of opinions in online discussions in form of

linguistic statements. Huang, Liu and Wang (2007)

introduce a method for detecting and tracking the

evolution of online communities. Choudhury, Sun-

daram, John, and Seligmann (2009) extract and

monitor key groups in blogs in order to study the

dynamics of the whole community. These approach-

es deal with the dynamic evolution in the past.

Other approaches take online chatter as a basis

for prediction. Gruhl, Guha, Kumar, Novak and

Tomkins (2005) use online postings to predict

changes and peaks in Amazon’s sales. Dahr and

Chang (2007) detect that user-generated content

correlates with future music sales. Onishi and Man-

chada (2009) arrive at the conclusion that blogging

activity correlates with the sales of green tea, movie

tickets and cell phone contracts. All three approach-

es do not predict the future behavior of Internet users

but only the consequences of users’ online activities.

There are also studies dealing with the prediction

of online behavior. Choudhury, Sundaram, John and

Seligmann (2007) describe a method for predicting

the communication flow in social networks. The

work of Choudhury (2009) allows the modeling and

forecasting of activities in online groups. However,

they do not identify opinion leaders and predict

opinion formation.

Welser, Gleave, Fisher and Smith (2007), Chang,

Chen and Chuang (2002) as well as Gomez, Kalten-

brunner and Lopez (2008) study social networks

with regard to the different roles of their members

(e.g. opinion leaders). However, the content of the

conversation is not taken into consideration. Boden-

dorf and Kaiser (2009) extract opinions by text min-

ing and identify opinion leaders with the aid of so-

cial network analysis in order to analyze opinion

formation.

None of the mentioned approaches focus on the

evaluation of the situation as a whole and none of

them consider all the influencing variables. Hence,

former work did not enable the recognition of criti-

cal situations and the alerting of marketing manag-

ers.

3 APPROACH

The objective of the approach is to detect critical

situations during opinion formation in online social

networks. Situations are considered as critical if

negative opinions are on the verge of spreading and

harming the company’s image or sales volume. In

these cases, marketing managers must be warned

immediately in order to take counteracting measures.

The approach comprises three succeeding mining

steps. In the first step, the opinions of all network

members towards a product are identified by me-

thods coming from text mining. Opinions are distin-

guished as positive, negative and neutral. In the

second step, the opinion leaders and the network

structure which have a great impact on the spread of

opinions are determined by using key figures from

social network analysis. In the third step, rules for

discovering critical situations during opinion forma-

tion are revealed on the basis of the overall opinion

of the network, the opinion and power of the opinion

leaders as well as the network structure. With the aid

of a fuzzy perceptron, linguistic rules are learned

which can be easily understood by marketing man-

agers. Rules learned from past situations can be

employed to recognize future critical situations and

to warn marketing managers at an early stage, i.e.

before the spread of negative opinions.

4 DATA COLLECTION

The presented approach is applied to the German

Gaming community Gamestar.de for purposes of

illustration and validation. The online platform Ga-

mestar.de is provided by Europe’s most popular

magazine for computer games. Fans of computer

games meet frequently on Gamestar.de to exchange

opinions on many games within the discussion fo-

rum. 6596 postings submitted from Oct. 8

to Nov.

2008 were extracted from threads discussing the

games “Fallout 3”, “Far Cry 2” and “Dead Space”.

For each of these three games, a sequence of time-

dependent networks was generated by connecting

those people with each other who have submitted

postings directly before or after one another on one

day.

DISCOVERING CRITICAL SITUATIONS IN ONLINE SOCIAL NETWORKS - A Neuro Fuzzy Approach to Alert

Marketing Managers

5 IDENTIFICATION OF

OPINIONS

The identification of opinions aims at detecting the

attitude of each user towards a product on the basis

of his/her postings. Attitudes are assigned according

to their polarity to the classes “positive”, “negative”

or “neutral”.

The process of opinion formation consists of two

phases (Kaiser and Bodendorf 2009). First, the post-

ings of the forum users are characterized by

attributes. Second, the postings are classified accord-

ing to their polarity on the basis of these attributes.

Statistical and linguistic attributes are used to de-

scribe the postings. For this reason, postings are

decomposed into words. Unimportant stop words are

removed. All remaining words are reduced to their

word stem. The relative frequency of each word

stem for each of the three classes is then calculated.

Word stems which appear frequently in one class but

rarely in the other two classes are chosen as

attributes.

Several methods such as Hidden Marcov Models

or Maximum Entropy enable the solving of classifi-

cation tasks (Weiss 2005). Support Vector Machines

(Cortes and Vapnik 1995) are specially suited for

text classification since they are able to process

numerous attributes. A lot of papers (e.g. Pang et al.

2002) have empirically demonstrated the appro-

priateness of Support Vector Machines for text clas-

sification. Therefore, Support Vector Machines are

employed for classifying the polarity of postings

based on their attributes.

In order to learn classification, training data con-

sisting of the postings’ attributes and manually as-

signed polarities are required. With the aid of this

training data, Support Vector Machines learn the

parameters of binary classification rules. In the case

of three classes, three classification rules are

learned: “positive” versus “not positive”, “negative”

versus “not negative” and “neutral” versus “not

neutral”. The final decision to which class a posting

is assigned is based on a majority vote. In the simple

case of just two attributes, the classification rule can

be depicted as a straight line separating the postings

into two classes. Figure 1 shows a line which classi-

fies a posting as positive due to the word stems it

contains.

After classification, the average opinion of each

user is determined based on all the postings he/she

has submitted to the discussion forum per day

In order to validate this procedure, it was applied

to the German Online Forum of Gamestar.de. 4010

postings from threads discussing the games “Dead

contains „love“

contains „best“

I love this game! It is the

best game I‘ve ever played.

Figure 1: Classification of opinions.

Space”, “Fallout 3” and “Far Cry 2” were manually

classified as positive, negative and neutral. The

validation is executed in form of a stratified 10-fold

cross-validation. The data set is divided into 10

portions so that each portion contains the same

amount of postings per class. Each of the portions is

used once to test the rules learned on the basis of the

other nine portions. After all ten test runs, the aver-

age performance in form of precision, recall and F-

measure is calculated (Weiss, Indurkhya, Zhang and

Damerau 2005). While precision measures the accu-

racy of the classification learned, recall determines

its completeness. The F-measure is calculated as the

harmonic mean of precision and recall.

Table 1 shows the results of the cross-validation.

While the detection of negative and neutral opinions

is very good, the detection of positive opinions is

less successful. The examination of misclassification

reveals that those postings with a positive introduc-

tion or positive conclusion but a neutral or negative

statement in-between are often not recognized as

positive. This problem can be solved by attaching

more weight to the sentences at the beginning and

the end of a posting.

Table 1: Results of opinion classification.

Class Precision Recall F-Measure

positive 62,96% 62,52% 62,74%

negative 86,35% 86,05% 86,20%

neutral 81,65% 81,07% 81,36%

6 CHARACTERIZATION OF

SOCIAL NETWORK

6.1 Opinion Leaders

Opinion leaders are persons who have great influ-

ence on other people’s opinion, attitude and beha-

vior (Katz and Lazarsfeld 1955; Rogers 2003).

While prior research in social psychology characte-

rized opinion leaders on the basis of their personal

KDIR 2010 - International Conference on Knowledge Discovery and Information Retrieval

attributes such as age and education, recent research

defines opinion leaders on the basis of their social

activities. Due to their central position and commu-

nicative behavior they play a leading role in opinion

formation (Valente 1999). A small number of opi-

nion leaders is sufficient to influence the opinions of

many others in a network (Keller and Berry 2003).

Persons are not split up in the classes “opinion lead-

er” or “no opinion leader” but are characterized by

the degree to which they affect others’ opinions

(Rogers 2003).

Social Network Analysis provides three key fig-

ures for measuring the degree of opinion leadership:

degree centrality, closeness centrality and between-

ness centrality (Wassermann and Faust 1999, Scott

2000). The normalized values of these centrality key

figures range from zero to one. While a value of one

indicates maximum opinion leadership, a value of

zero indicates minimum opinion leadership.

Degree centrality measures how many direct re-

lationships a person has to other network members.

It is calculated as the ratio of the number of a user’s

relationships to the number of all relationships in the

network. Degree centrality specifies how often a

person communicates directly with other persons in

the network. Persons with high degree centrality are

in a position to influence their local surroundings

and can be considered as local opinion leaders.

In contrast to degree centrality, closeness cen-

trality does not only take direct but also indirect

communication relationships into account. Closeness

centrality characterizes how close a person is to all

other persons in the network. It is calculated as the

inverse sum of the distances from each user to all

other users within the network. Persons with high

closeness centrality are able to influence the overall

network due to their short distance to all other users

in the network. Therefore, they can be considered as

global opinion leaders.

Betweenness centrality describes how frequently

a user can be found on the shortest connecting paths

between all pairs of users. This key figure is deter-

mined as the fraction of the shortest paths which

pass a user to all shortest paths within the network.

Since a lot of communication flows via persons with

high betweenness centrality they act as intermedia-

ries and have the power of influencing the flow of

information.

6.2 Network Structure

Besides opinion leaders, the structure of the network

has an impact on opinion formation. The network

structure can be characterized by the key figures

centralization and density coming from the social

network analysis (Wassermann and Faust 1999,

Scott 2000).

Centralization measures how the centralities of

the network members differ from the centrality of

the most central person. A strongly centralized net-

work consists of only a few central opinion leaders

and many peripheral users. In this case, the leaders’

opinion can spread easily from the center to the

periphery of the network (Bodendorf and Kaiser

2009).

Density specifies the connectivity of a network.

It is calculated as the fraction of the number of rela-

tionships which exist in a network and the maximum

number of relationships which are possible in a

network. Density indicates the frequency of commu-

nication within the network. The higher the density

of a network, the more opinions can be exchanged

between the network members (Bodendorf and

Kaiser 2009). In a very dense network the opinion

can disseminate quickly among the network mem-

bers.

7 DISCOVERY OF CRITICAL

SITUATIONS

7.1 Objective

The objective is to discover critical situations auto-

matically and to alert marketing managers as soon as

such a situation arises in the process of opinion

formation. The classification of situations depends

on many different variables such as the overall opi-

nion, the opinions of the opinion leaders or the struc-

ture of the network.

The approach attempts to fulfill two countervail-

ing requirements. On the one hand, marketing man-

agers should be in a position to easily comprehend

why a situation is classified as critical. Consequent-

ly, the system must be able to process linguistic

rules that can be formulated by the managers due to

their expertise. On the other hand, interdependencies

among influencing factors are very complex and

make it difficult for marketing managers to define

rules for detecting critical situations. For this reason,

the system must enable supervised learning of such

linguistic rules from data.

7.2 Method

With regard to the two requirements mentioned

above, two methods coming from the discipline of

DISCOVERING CRITICAL SITUATIONS IN ONLINE SOCIAL NETWORKS - A Neuro Fuzzy Approach to Alert

Marketing Managers

soft computing come into consideration, i.e. artificial

neural networks and fuzzy systems.

Neural networks are capable of learning classifi-

cation from data. However, they are a black box.

There is no way of understanding the rules behind

the classifications (Nauck, Klawonn and Kruse

1997). In contrast, fuzzy systems cannot learn classi-

fication from data but they can process linguistic

rules that are based on fuzzy sets (Zadeh 1965).

Experts formulate such linguistic rules which are

then employed for classification. This enables the

understanding of classification results.

In order to combine the advantages and minimize

the disadvantages of neuronal networks and fuzzy

systems (Nauck et al. 1997), a neuro fuzzy approach

is applied in this work.

Neuro fuzzy systems have the ability of learning

linguistic rules from data. There are many different

neuro fuzzy approaches. Here the NEFCLASS mod-

el (NEuro Fuzzy CLASSification) is chosen. This

system is capable of learning fuzzy sets and fuzzy

rules. Moreover, it can also deal with manually de-

fined rules and optimize them (Nauck et al. 1997).

The NEFCLASS model is a 3-layer fuzzy per-

ceptron (Nauck and Kruse 1994). The input layer

represents the input variables, the hidden layer the

fuzzy rules and the output layer the two classes (crit-

ical situation and non-critical situation). The linguis-

tic terms are represented by the weights between the

input layer and the hidden layer (Nauck and Kruse

1995). Figure 2 illustrates the structure of a fuzzy

perceptron.

critical

situation

non‐

critical

situation

rule 1 rule 3 rule 4

overall

opinion

opinion of

the opinion

leader

...

rule 2

linguistic

terms

alert nonalert

neutral

negative

positive

negative

neutral

positive

…

rules

class

output

input

Figure 2: Structure of a fuzzy perceptron.

The fuzzy perceptron depicted in figure 2 con-

sists of four fuzzy rules. For example, rule 1 classi-

fies a situation as critical if the input variables over-

all opinion and opinion of opinion leader take the

value of the linguistic term negative (see figure 3).

Rule1



Ifoverallopinionisnegative

andopinionofopinionleaderisnegative

thensituationiscritical

Figure 3: Fuzzy rule.

Fuzzy sets specify whether and to what degree the

values of the input variables belong to linguistic

terms. While in classical set theory objects either

belong or do not belong to a set, in fuzzy set theory

objects belong to a set with a certain degree of

membership. A fuzzy set is a function which assigns

a degree of membership for a linguistic term to each

value of the input variable (see figure 4).

negativeneutralpositive

Opinion

Degreeof Membership

Figure 4: Fuzzy sets.

The fuzzy rules are learned from data by employing

an algorithm derived from the algorithm of Wang

and Mendel (Wang and Mendel 1991 & 1992, Bor-

gelt, Klawonn, Kruse and Nauck 2003). The feature

space is structured by overlapping hyperboxes which

represent fuzzy rules (Nauck et al. 1997). Each

hyperbox is an n-dimensional Cartesian product of n

fuzzy sets (Nauck and Kruse 1997). Figure 5 shows

a feature space that is structured by overlapping

hyperboxes. In this case, there are two variables:

opinion leadership (with the fuzzy sets low, medium,

high) and opinion (with the fuzzy sets negative,

neutral, positive).

The algorithm learns the fuzzy rules by running

through the training data set twice. In the first run,

all antecedents (if parts) of the rules are generated.

For each pattern of the training data set, the combi-

nation of fuzzy sets which achieves the highest de-

gree of membership is selected. In the second run,

the rules are completed by determining the best

consequent (then part) for each antecedent (if part).

The resulting rules enable the classification of input

patterns. However, there may still be some classifi-

cation errors. Figure 5 (left side) exemplifies the

classification of the input pattern (circles and trian-

gles) after rule learning. One pattern is not classified

(triangle) and one is misclassified (circle).

KDIR 2010 - International Conference on Knowledge Discovery and Information Retrieval

opinion

positiveneutralnegative

opinion

leadership

low medium high

criticalsituations

non‐criticalsituations

opinion

positiveneutralnegative

opinion

leadership

low medium high

Figure 5: Classification after rule learning (left side) and after fuzzy set modification (right side).

On the basis of the detected errors the shape and

the position of the fuzzy sets are modified in order to

improve classification (Nauck and Kruse 1995).

Figure 5 (right side) shows the classification results

after the process of modification. All patterns are

classified. There are no misclassifications.

After modification of the fuzzy sets, the rule base

is pruned by deleting variables or whole rules. Thus,

the rule base is easier to interpret and can be applied

to a broader range of cases (Nauck 1996).

7.3 Application

Training Sets

Since the NEFCLASS model is based on supervised

learning, the training datasets must be classified

manually before learning. Each day’s snapshot of

the discussion network is evaluated as a critical or

non-critical situation.

Figure 6 shows a situation which is not critical.

The local opinion leader and the intermediary is

User 25 (degree centrality 0.57, betweenness cen-

trality 0.48). The global opinion leader is User 26

(closeness centrality 0.65). The opinion of both of

them is positive. The overall opinion is neutral. The

centralization, i.e. the likelihood that the opinion of

the opinion leader will diffuse, is medium (0.4). The

density, i.e. the speed of diffusion, is small (0.18).

Consequently, this situation is classified as non-

critical.

‐

USER1

USER2

USER3

USER4

USER5

USER6

USER8

USER7

USER9

USER10

USER26:GLOBAL

OPINIONLEADER

USER11

USER12

USER3

USER13

USER15

USER14

USER16

USER17

USER18

USER19

USER20

USER23

USER21

USER22

USER27

USER24

USER25:LOCALOPINIONLEADER

&IN TERMEDIARY

+ positive opinion

‐ negativeopinion

else:neutralopinion

Figure 6: Non-critical situation in a network.

In contrast, the situation in figure 7 is classified

as critical. Local opinion leader and intermediary is

User 8. Global opinion leader is User 9. The opinion

of all opinion leaders is neutral. The overall opinion

is slightly negative (-0.1). The likelihood that the

opinion of the opinion leader will diffuse is high

(centralization 0.47). The speed of diffusion is me-

dium (density 0.36). The situation is critical since a

neutral opinion can be interpreted as disinterest.

DISCOVERING CRITICAL SITUATIONS IN ONLINE SOCIAL NETWORKS - A Neuro Fuzzy Approach to Alert

Marketing Managers

USER1

USER2

USER3

USER4

USER5

USER6

USER7

USER8:LOCALOPINION

LEADER&INTERMEDIARY

USER9:GL OBAL 

OPINIONLEADER

‐

+ positive opinion

‐ negativeopinion

else:neutralopinion

Figure 7: Critical situation in a network.

Rule Bases

Based on the training datasets, the classification

rules for all three games are learned. The resulting

rule bases for all three games are clear and easily

interpretable. They consists of eight rules for the

game “Dead Space”, nine rules for the game “Fal-

lout 3” and three rules for the game “Far Cry 2”.

Figure 8 shows an extraction of the rule base for the

game “Dead Space”.

Rule 1

If the local opinion leadership is high

and the opinion of the local opinion leader is negative

and the opinion of the global opinion leader is negative

and the likelihood that the opinion of an opinion leader

will diffuse is high

and the overall opinion is negative

then the situation is critical.

Rule 2

If the opinion of the global opinion leader is positive

and the likelihood that the opinion of an opinion leader

will diffuse is low

and the overall opinion is positive

then the situation is not critical.

Figure 8: Extraction of the rule base.

The first rule classifies critical situations. It states

that if the overall opinion is negative and the opinion

of the local and the global opinion leader is negative

as well, then the marketing manager must be alerted.

It is also important that local opinion leadership and

centralization, i.e. likelihood of opinion diffusion,

are high. In these situations the opinion will remain

negative in the future or even become more nega-

tive. As a consequence thereof marketing manager

must take actions to influence opinion formation.

The second rule classifies situations that are not

critical. The opinion of the global opinion leader is

positive. The probability that the opinion of the opi-

nion leader will diffuse is low. However, this is of

no disadvantage since the overall opinion is already

positive. There is no indication that the overall opi-

nion will become negative in the future. For this

reason the marketing manager there is no need for

altering the marketing manager.

Classification Results

Table 2 depicts the classification results of the three

games. It shows the average rate of misclassification

during validation. The best results have been

achieved for the game “Fallout 3”. The validated

classifier has an estimated misclassification rate of

6.5%.

The classifier learned for the game “Dead

Space” is also excellent. The average rate of mis-

classification is 9.7%. Due to the small training

dataset, learning is less successful for the game “Far

Cry2”. There is a misclassification rate of 29.4% for

the game “Far Cry 2”.

Table 2: Classification results.

Game

Number of time-

dependent networks

Misclassifi-

cation rate

Dead Space 56 9.7%

Fallout 3 45 6.5%

Far Cry 2 33 29.4 %

8 CONCLUSIONS

The presented approach enables the detection of

critical situations during opinion formation in online

social networks by executing three mining steps.

First, the opinions of all network members towards a

product are recognized by methods coming from text

mining. Second, the opinion leaders and the struc-

ture of the network are determined by key figures

KDIR 2010 - International Conference on Knowledge Discovery and Information Retrieval

coming from social network analysis. Third, critical

situations during opinion formation are spotted by a

fuzzy perceptron on the basis of the opinions of the

network members, the influence of the opinion lead-

ers as well as the structure of the network. Choosing

a neuro fuzzy approach allows the learning of lin-

guistic rules which can be easily interpreted by mar-

keting managers. These rules are learned from past

situations and can be employed to judge future situa-

tions.

There are a lot of advantages in discovering crit-

ical situations. Being alerted at an early stage, mar-

keting managers can influence the process of opi-

nion formation. For instance, they can address opi-

nion leaders who have a negative opinion and ask

their advice about product improvements. This ac-

tion might not only reveal valuable information for

product development but might also lead to a change

in the leaders’ opinions as they have the impression

that their complaint is being taken seriously. All in

all, this approach attempts to improve a company’s

image and to increase its sales volume.

Scheduled work is to implement a decision sup-

port system that not only identifies critical situations

but also generates recommendations on appropriate

actions for marketing mangers. For example, the

system should advise marketing managers how to

communicate with network members in critical

situations.

REFERENCES

Bodendorf, F., Kaiser, C., (2009). Detecting Opinion

Leaders and Trends in Online Social Networks. In

Proceedings of the 2nd Workshop on Social Web

Search and Mining. Hong Kong.

Borgelt, C., Klawonn, F., Kruse, R., Nauck, D., 2003.

Neuro-Fuzzy-Systeme: Von den Grundlagen

künstlicher Neuronaler Netze zur Kopplung mit Fuzzy

Systemen. [engl.: Neuro-Fuzzy-Systems: Foundations

of the combination of neural networks and fuzzy-

systems.] (3th ed.). Wiesbaden: Vieweg.

Chang, C. L., Chen, D. Y., and Chuang, T. R., (2002).

Browsing Newsgroups with a Social Network Analyz-

er. In Proceedings of the Sixth International Confe-

rence on Information Visualization, London.

Choudhury, M. D., Sundaram, H., John, A., Seligmann, D.

D. (2007): Contextual Prediction of Communication

Flow in Social Networks. In Proceedings of the

IEEE/WIC/ACM international Conference on Web in-

telligence (Silicon Valley, California, USA). Web In-

telligence. IEEE Computer Society, Washington, DC,

pp. 57-65.

Choudhury, M. D. (2009). Modelling and Predicting

Group Activity over Time in Online Social Media. In

Proceedings of the Twentieth ACM Conference on

Hypertext and Hypermedia. Torino, Italy.

Choudhury, M. D., Sundaram, H.,John, A., Seligmann, D.

D. (2009). Which are the Representatative Groups in a

Community? Extracting and Characterizing Key

Groups in Blogs. ACM Student Research Competition,

HyperText '09.

Cortes C., Vapnik V. N., 1995. Support Vector Networks.

In Machine Learning, Vol. 20, pp. 273-297.

Dave, K., Lawrence, S., Pennock, D., M. (2003). Mining

the peanut gallery: Opinion extraction and semantic

classification of product reviews. In Proceedings of

the 12th international conference on World Wide Web.

Dhar, V., Chang, E. (2007). Does Chatter Matter? The

Impact of User-Generated Content on Music Sales.

Technical Report, Leonard N. Stern School of Busi-

ness, New York University.

Glance, N., Hurst, M., Nigam, K., Siegler, M., Stockton,

R., Tomokiyo, T. (2005). Deriving Marketing Intelli-

gence from Online Discussion. In Proceedings of the

eleventh ACM SIGKDD international conference on

Knowledge discovery in data mining. Chicago, Illi-

nois, USA, pp. 419 – 428.

Gomez, V., Kaltenbrunner, A., and Lopez, V. (2008).

Statistical Analysis of the Social Network and Discus-

sion Threads in Slashdot, In Proceedings of the Inter-

national World Wide Web Conference, Beijing: ACM

Press.

Gruhl, D., Guha, R., Kumar, R., Novak, J., Tomkins, A.

(2005). The Predictive Power of Online Chatter. In

Proceedings of the eleventh ACM SIGKDD interna-

tional conference on Knowledge discovery in data

mining. Chicago, Illinois, USA, pp. 78 - 87.

Huang, Y., Liu, S., Wang, Y. (2007). Online Detecting

and Tracking of the Evolution of User Communities.

In Third International Conference on Natural Compu-

tation, pp.681-685.

Kaiser, C. (2009): Combining Text Mining and Data

Mining For Gaining Valuable Knowledge from Online

Reviews. In Pedro Isaías (Ed.). IADIS International

Journal on WWW/Internet 6 (2), pp. 63-78, 2009.

Kaiser, C., Bodendorf, F. (2009). Opinion and Relation-

ship Mining in Online Forums. In Proceedings of the

2009 IEEE/WIC/ACM International Joint Conference

on Web Intelligence and Intelligent Agent Technology.

Milan, pp. 128-131.

Katz E., Lazarsfeld P. F, (1955). Personal influence, the

part played by people in the flow of mass communica-

tion, Glencoe: Free Press.

Keller E. B., Berry J., (2003). The influentials, New York:

Free Press.

Kim, S.-M., Hovy, E. (2007). Crystal: Analysing Predic-

tive Opinions on the Web. In proceedings of the 2007

Joint Conference on the Empirical Methods in Natural

Language Processing and Computational Natural

Language Learning, Prague, pp. 1056-1064.

Nauck, D., Kruse, R., (1994). A Fuzzy Perceptron as a

Generic Model for Neuro-Fuzzy Approaches. In

Fuzzy Systeme '94.

DISCOVERING CRITICAL SITUATIONS IN ONLINE SOCIAL NETWORKS - A Neuro Fuzzy Approach to Alert

Marketing Managers

Nauck, D., Kruse, R., (1995). NEFCLASS - A Neuro-

Fuzzy Approach for the Classification of Data. In

George, K.M., Carrol, J. H., Deaton, E., Oppenheim,

D. Hightower, J. (Ed.), (1995). Applied Computing

1995: Proc. of the 1995 ACM Symposium on Applied

Computing, Nashville, Feb. pp. 26-28. ACM Press.

Nauck, D., Klawonn, F., Kruse, R., (1997). Foundations of

neuro-fuzzy systems. Chichester: John-Wiley & Sons.

Nauck, D., Kruse, R., (1997). A neuro-fuzzy method to

learn fuzzy classification rules from data. In Fuzzy

Sets and Systems 1997 (89), pp. 277-288.

Nauck, U., (1999). Design and Implementation of a Neu-

ro-Fuzzy Data Analysis Tool in Java. Diploma Thesis,

University of Braunschweig, Braunschweig.

Onishi, H., Manchanda, P. (2009): Marketing Activity,

Blogging and Sales. Technical Report, Ross School of

Business, University of Michigan.

Pang P., Lee L., and Vaithyanathan S., (2002). Thumbs

up? Sentiment Classification using Machine Learning

Techniques. In Proceedings of the conference on em-

pirical methods in natural language processing.

ACM, pp. 79-86.

Popescu, A.-M., Etzioni, O. (2005). Extracting Product

Features and Opinions from Reviews. In Proceedings

of Human Language Technology Conference and Con-

ference on Empirical Methods in Natural Language

Processing (HLT/EMNLP), pp. 339-346.

Rogers E., (2003). Diffusion of innovations (5

ed.). New

York: Free Press.

Scott J., (2000). Social Network Analysis – A Handbook.

London: SAGE.

Tong, R. M., Yager, R. R. (2004). Characterizing Attitu-

dinal Behaviors in On-Line Open-Source. Proceedings

of Association for the Advancement of Artificial Intel-

ligence. Spring Symposium 2004, Atlanta.

Valente T. W., (1999). Network Models of the Diffusion of

Innovations, Cresskill: Hampton Press.

Viermetz, M., Skubacz, M., Ziegler, C.-N.; Seipel, D.

(2005). Tracking Topic Evolution in News Environ-

ments. In 10

IEE Conference on E-commerce Tech-

nology and the Fifth IEEE Conference on Enterprise

Computing, E-Commerce and E-Services, pp. 215-

220.

Wang, L.-X., Mendel, J.M., (1991). Generating Rules by

Learning from Examples. In Int. Symposium on Intel-

ligent Control. Piscataway, NJ, USA: IEEE Press,

pp.263-268.

Wang, L.-X., Mendel, J.M. (1992). Generating Fuzzy

Rules by Learning from Examples. IEEE Trans. Sys-

tems, Man, and Cybernetics, 22 (6), pp. 1414-1427.

Piscataway, NJ, USA: IEEE Press.

Wassermann, S., Faust, K., (1999). Social Network Analy-

sis – Methods and Applications. Cambridge: Cam-

bridge University Press.

Weiss S., Indurkhya N., Zhang T., Damerau F. (2005).

Text Mining – Predictive Methods for Analyzing un-

structured Information. New York: Springer.

Welser, H. T., Gleave, E., Fisher, D., Smith, M. (2007).

Visualizing the Signatures of Social Roles in Online

Discussion Groups. In Journal of Social Structure,

Vol. 8.

Zadeh, L., (1965). Fuzzy Sets. In Information and Con-

trol 8 1965 (3), San Diego, CA, USA: Academic

Press, pp. 338-353.

KDIR 2010 - International Conference on Knowledge Discovery and Information Retrieval