Portmann 2015

JID:FSS AID:6570 /FLA [m3SC+; v 1.194; Prn:7/07/2014; 16:20] P.
1 (1-25)
Available online at www.sciencedirect.com
ScienceDirect
Fuzzy Sets and Systems ••• (••••) •••–•••
www.elsevier.com/locate/fss
FORA – A fuzzy set based framework for online reputation

management
Edy Portmann a,b,∗ , Andreas Meier c , Philippe Cudré-Mauroux c , Witold Pedrycz d,e,f
a University of Bern, Engehaldenstrasse 8, 3012 Bern, Switzerland
b University of California, Berkeley, 443 Soda Hall, Berkeley, CA 94720, USA
c University of Fribourg, Boulevard de Pérolles 90, 1700 Fribourg, Switzerland
d Department of Electrical & Computer Engineering, University of Alberta, Edmonton, T6R 2V4 AB, Canada
e Department of Electrical and Computer Engineering, Faculty of Engineering, King Abdulaziz University, Jeddah 21589, Saudi Arabia
f Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland
Received 19 February 2013; received in revised form 25 March 2014; accepted 4 June 2014
Abstract
The Social Web offers increasingly simple ways to publish and disseminate personal or opinionated information, which can
rapidly exhibit a disastrous influence on the online reputation of organizations. Based on social Web data, this study describes the
building of an ontology based on fuzzy sets. At the end of a recurring harvesting of folksonomies by Web agents, the aggregated
tags are purified, linked, and transformed to a so-called fuzzy grassroots ontology by means of a fuzzy clustering algorithm. This
self-updating ontology is used for online reputation analysis, a crucial task of reputation management, with the goal to follow the
online conversation going on around an organization to discover and monitor its reputation. In addition, an application of the Fuzzy
Online Reputation Analysis (FORA) framework, lessons learned, and potential extensions are discussed in this article.
© 2014 Elsevier B.V. All rights reserved.
Keywords: Fuzzy clustering; Grassroots ontology; Online reputation management; Search engine; Social semantic web; Topic Maps
1. Motivation
An online reputation is the publicly kept social evaluation of an entity based on the entities previous behavior, what
was posted by the entity, and what third parties share about the entity on the Internet. The Internet thereby constitutes a
worldwide database, where information is archived and not easily deleted. People, organizations, and governments are
increasingly drawing on today’s social Web to create and share on the Internet. In large parts, this Web is comprised
of media that are simultaneously hugely scalable and easy-to-use. It provides prosumers (combination of producer
and consumer of Web data) publishing techniques to spread information through social interaction. One of its key
* Corresponding author.
E-mail addresses: edy.portmann@iwi.unibe.ch (E. Portmann), andreas.meier@unifr.ch (A. Meier), philippe.cudre-mauroux@unifr.ch
(P. Cudré-Mauroux), wpedrycz@ualberta.ca (W. Pedrycz).
http://dx.doi.org/10.1016/j.fss.2014.06.004
0165-0114/© 2014 Elsevier B.V. All rights reserved.
JID:FSS AID:6570 /FLA [m3SC+; v 1.194; Prn:7/07/2014; 16:20] P.2 (1-25)
2 E. Portmann et al. / Fuzzy Sets and Systems ••• (••••) •••–•••
Fig. 1. Concept of knowledge map and its visualization.
characteristics is that it can stimulate to echo information; as a consequence opinions are easily shared through the
proliferation of blogs, microblogs, social networks and other social media sites [44]. Nothing can ever again avoid
an organization’s danger from being the focus of a negative entry [10]. By this means social media can severely
affect the public perception of an organization. Online buzz may be good or bad, and consequently a Web strategy
for a first-hand interaction with stakeholders to respond to their concerns, is absolutely invaluable [34]. It is vital for
an organization to listen to what is being said about it, even (or not until then) if it is negative. Listening bears the
opportunity to take action and resolve problems or deal with virulent information, both of which can negatively affect
an organization’s brand image and reputation.
To not degenerate into an organization that is indeed social Web data rich but knowledge poor [1,31], for on-
line reputation management a straightforward graphical user interaction is strived for, which helps organizations to
uncover contexts. Through an interactive visualization of an automatically generated knowledge base as associative
nets, a responsible communication operative can easily grasp semantic context and straightforwardly communicate
with potential attackers of the organizations reputation. Accordingly, this article presents an ontology-based knowl-
edge map system to properly organize Web data into Topic Maps [32], from which a communication operative can
obtain in-depth concepts to facilitate addressing critics and eventually go into an appropriate online conversation.
A Topic Map thereby characterizes a standard for representing knowledge structures and associating information and
resources to those structures. It represents the topics, associations (i.e., the relationships between the topics), and oc-
currences (i.e., information that is considered relevant to the topic). Since Topic Maps are not as suitable for general
axiomatic specifications as the description logic based ontologies (i.e., a formal, explicit specification of a shared con-
ceptualization [14,15]), in this article, they are used as visualization technique. Yet, in the Semantic Web ontologies
are built on Boolean (two-valued) logic that is insufficient to reflect social Web’s reality. Fig. 1 illustrates the concept
of knowledge maps and its visualization as Topic Maps a communication operative can interact.
Fig. 1a illustrates the knowledge base that besides the topics and associations also includes occurrences (i.e., terms
with associated Web data). The incidental terms are Web data specific and relevant to a particular topic. A good com-
bination of map validity, novelty, usefulness, simplicity, understandability, and generality thereby is a main objective
during the creation of the framework [14,31]. Hence, the visualization of the map (see Fig. 1b) can help users to locate
required information and also offers subject-related information easily and rapidly [1].
In this study, we present the fuzzy-based framework FORA for online reputation analysis and management. Using
this framework, an organization can independently follow the online conversation going on around its brand, com-
petitor’s brand or discover and monitor real-time reputation of a given business sector. Conversations in the social
Web rely on the use of natural language that convey a complex structure, which is substantive not only to the human
communication, but also to the way human beings think and perceive the world. The main idea and the underlying
model of our approach is to capture the imprecision of human language used in Web content, and to express it with
appropriate mathematical tools for the realization of the framework. By fuzzy set theory human vagueness is cap-
tured and expressed in a formal fashion [29,42,47]. Accordingly, fuzzy set theory provides a basis for an emulation
of higher order cognitive functions, thoughts and perceptions inherent in social Web data. For this very reason the
E. Portmann et al. / Fuzzy Sets and Systems ••• (••••) •••–••• 3
framework is based on fuzzy set theory to handle such vague data in an easy and convenient fashion. The use of fuzzy
sets allows organically grooming gradual knowledge structures from existing Web content. The lower a membership
degree for belonging to a topic, the farther away it should be represented within the Topic Map [23,24]. A fuzzy
membership degree thereby expresses the Social Web’s different knowledge granularity that is automatically captured
by the framework.
In the present business landscape, Sen [39] roughly divides online reputation analysis and management applica-
tions into technical and intellectual models: the first abstain largely from human factors (e.g., manual input, human
knowledge, interpretation capability, etc.) and are mainly limited to specific use of purpose-built search engines (e.g.,
Actionly, MeMo, Sysomos, etc.; for a comparison see Subsection 4.3). With their powerful visualization techniques
as primary advantage, however, they are the most widespread these days. Pure intellectual models, on the other hand,
are by contrast only just based on human factors to find and evaluate social Web data. Yet, in its quality and complex-
ity, human perception-based intellectual models are in the ascendancy over fully automated technical ones. On that
account, the intention of our approach is to combine the strengths of both models and thereby to allow a generally
automated but human-centered online reputation analysis as suggested in [1,31]. In this article we try to combine not
only public relations with information technology topics, but also practical with theoretical ones. Such balancing acts
are always out on a limb for research from business informatics [18]. As biggest pitfall, the act thereby entails the
difficulty to bring everything under one roof.
Section 2 is devoted to a brief introduction into online reputation analysis, which constitutes a key challenge of
reputation management. On this basis, Section 3 provides the Fuzzy Online Reputation Analysis (FORA) framework.
The term FORA stems from the plural form of forum, the Latin word for marketplace. The framework allows a fuzzy
exploration of reputation on online marketplaces. Section 4 introduces the YouReputation prototype as instantiation of
the framework. Thereby the term YouReputation is a portmanteau formed by contracting the word your with the word
reputation, voicing the importance of an organization’s online reputation management. In this section also lessons
learned are presented. Finally, conclusions and future research directions are outlined in Section 5.
2. Online reputation analysis
Oscar Wilde once observed that “one can survive everything, nowadays, except death, and live down everything
except a good reputation” [25]. Hence, reputation is an important, albeit intangible asset, which can cause organi-
zations to fail or succeed. Sticks and stones may break bones but a bad reputation can take an organization out of
business. Thereby, an organization’s reputation comes across as a social evaluation of stakeholders toward the man-
agement of different criteria or, metaphorically speaking, as the result of what an organization does, says, and what
consumers say about the organization [10]. Although an organization’s reputation is built upon trust [44], in turn,
trust is an outcome of a sound reputation – these concepts mutually influence each other [10]. Yet, organizations have
repeatedly asked customers to trust them. In a prosumer-oriented market as on the Web, the tide has turned and organi-
zations themselves must show that they trust the consumers. “Online is where reputations are made now” according to
Leslie Gaines-Ross (in [4]), and so it is reasonable for an organization to implement an online reputation management
strategy, because whether an organization trusts the consumers or not, they will talk online about the organization in
reputation-increasing or reputation-detracting mode. Listening to the online chatter will give an organization insight
about its reputation.
The next subsection clarifies the importance of measuring reputation capital (see Subsection 2.1). This is followed
by an introduction into online reputation management process, which pays extra tribute to an organizations reputation
capital in Subsection 2.2. In Subsection 2.3, concluding, we expound why applying fuzzy set theory to vague Social
Web works very well.
2.1. Measuring the reputation capital
Reputation capital refers to a crucial but sometimes under-evaluated asset of an organization that can be managed,
accumulated and traded in for trust, legitimization (e.g., of a position of power and social recognition), premium
prices (e.g., for goods and services), greater readiness among shareholders to hold on to shares in times of crisis, or
a stronger willingness to invest in the organization’s shares [25]. Providing functional and social expectancies of the
public for one thing and manage to build a remarkable identity for another originates trust that causes an informal
Fig. 2. Online reputation management process.
framework that, in turn, provides return in cooperation and yields reputation capital. A positive reputation ensures an
organization’s long-time competitive edges. The greater the reputation capital the less the costs for monitoring and
controlling may be [25]. Victor et al. present an aggregated trust and distrust score which might just be called in for
measuring reputation capital in particular (see [44]).
Anyway, as a consequence, reputation capital should be a key measure of an organization. In a world, where
the Web is becoming more and more social, to measure the effects of collaboration and contribution to an online
community is becoming more and more valuable too. Thus, in the Social Web it is viewed as some kind of non-cash
remuneration for organizations’ efforts, and usually generates respect with prosumers of an online marketplace where
capital is generated. Some prominent examples for measuring reputation capital in the Web are eBay’s seller rating
mechanism, Google’s PageRank, Technorati’s authority index or the Klout score. Online reputation capital, however,
is not only interesting for organizations to measure, but also for private persons (e.g., when someone is chasing a
dream school, job or relationship, the first impression most often is made online [36]).
2.2. The process of online reputation management
Online reputation management is the task of monitoring, addressing, or rectifying undesirable or negative mentions
on the Web. A corresponding strategy corollary monitors online buzz and sentiment about the organization by engag-
ing stakeholders positively. Such a strategy can roughly be divided into two steps: the first aims at finding important
online buzz and the second is about conveying an appropriate message to maintain a positive online sentiment. This
is illustrated in Fig. 2. The conveying of an online message entails an immediate interactive communication towards
stakeholders, whereas online reputation analysis proves necessary as an instrument to measure the social Web’s data
on the organization’s reputation. In the following, the latter is addressed that involves scanning, monitoring, and fore-
casting reputation-related issues [34]. Note, however, that a closed online reputation management strategy also entails
some continuous control [4,23,25].
The aim of scanning is the early detection of changes in the environment of the organization that may affect or
restrict its scope. Moreover, it allows spotting new sectors, which the organization can occupy to position itself as an
expert and opinion leader and corollary to realize new opportunities. Another goal of this approach is to evaluate the
reputations of competitors; occasionally, a competitor will launch an unknown product or a new production method
that can be detected through online scanning [34]. A challenge of reputation scanning is the prevention of flooding
caused by vast amounts of Web data. Issues must be summarized into manageable topics [1], and their changes
must be surveyed in the ensuing permanent monitoring to avoid any surprises. Monitoring is a method of reputation
analysis that is equivalent to scanning but watches a selected range of topics only. In addition, to reduce risks and
uncertainties, an organization can use forecasting that allows making assertions about reputation-related events whose
actual outcomes have not yet been observed. Forecasting can be used even before monitoring, and they may also
alternate [34].
2.3. Applying fuzzy sets in online reputation management
In today’s social Web, for organizations, issues with reputation may grow rather quick. Thereby, the degree of
public awareness defines its life cycle [21] whose time-related dynamics can be obtained from Fig. 3. In the beginning
an organizations scope of action is high, dealing with a certain online reputation issue is not too costly, and the public
attention is rather limited. But the more the public awareness increases, the less the organization can do in order to
solve the problem the easy way.
If a reputation issue escapes from online to offline (e.g., to the front page of the newspaper), it may have reached its
peak. That makes it easy to discover the issue; however, it may be too late for the organization to influences its progress
Fig. 3. Reputation life cycle diagram.
positive. To this end, the organization should have spotted the issue much earlier. Yet, since only few prosumers in the
respective online community were involved and the media did also not yet report about it, the issue was much harder
to discover. The evolution of a reputation issue is hard to foresee, a dichotomous phenomenon termed weak signal
problem [21]. This problem circumscribes the fact that it is often difficult to find new reputation issues as long as only
a few unknown stakeholders are involved [45].
Compared to traditional sets, variables from fuzzy set can have different memberships to different classes [48].
Fuzzy sets are graduated in the sense that membership in a class is a matter of degree. A fuzzy set A, in a universe
of discourse Ω, is defined by a membership function that associates with each object x in Ω, the degree to which x
is a member of A. A fuzzy set is basic if its membership function takes values in a unit interval [0, 1]. Applying the
concepts of fuzzy sets to the weak signal problem, allows finding reputation issues with the first mention in the Web.
A certain issue, that is to say, can more or less belong to a certain set; through co-occurrence on the same Web page,
an online reputation issue (i.e., an object x) that is connected with the organizations (i.e., a particular fuzzy set A such
as the organization’s name) can belong to this set from its first common mention. Yet, it can also belong to another
particular fuzzy set (e.g., the organization’s products or services). Since fuzzy set theory has been extended to handle
the concept of partial membership, where a membership value may range between completely true and completely
false, the motion along the partial membership to a certain class can help handling the weak signal problem. Note,
however, that in this article fuzzy reputation means something else as in the P2P domain.
Long story short, today’s organizations are compelled to listen to the Social Web so as to take part in and, in this
way, improve their online reputation [10]. To do this self-consciously, the FORA framework will be introduced.
3. The FORA framework
Folksonomies are typically based on vague human perceptions such as identity, location, status, time, and other
characteristic of physical and mental objects. A folksonomy is a Social Web classification practice and method of
collaboratively creating and managing tags to annotate and categorize content. In order to process this human in-
put, the FORA framework must be given the ability to process words, perceptions, and propositions drawn from
natural language. To this end, Zadeh’s computational theory of perceptions takes into account vagueness of human
concepts expressed using natural language [49]. Since fuzzy set theory provides a reasonable mathematical model of
the vagueness phenomenon, it is especially well suited for handling the spontaneously arising grassroots structures of
social Web data such as folksonomies – and fortunately also other social media this article addresses. In addition to
folksonomies, an ontology’s goal (see Section 1) is to afford meaning. In a semiotic sense, to obtain semantic, thereby
it is necessary to focus on the relation between signifier and their denota [8]. In contrast to a standardized ontology,
this insight based fuzzy grassroots ontology is thus useful for wielding natural bottom-up structures and therefore
to automatically add vague human semantics to the Social Web. In the end, our fuzzy grassroots ontology is about
identifying and formalizing a conceptualization of the activity of tagging that commits to the ontology at a semantic
level [15] using fuzzy set theory to determine gradations of tagged occurrences in a tagspace (i.e., a normalized and
linked folksonomy [15]). In doing so, set membership may be defined as a distribution f : [0, 1]χ → [0, 1], where χ
denotes conceivable possibilities.
Fig. 4. The FORA framework architecture.
The frameworks architecture is presented in Subsection 3.1. Then all layers and associated components are ex-
plained in more detail. In doing so, the construction of the particular components of the FORA framework is
emphasized. Subsection 3.2 starts with the reputation search engine, and then Subsection 3.3 introduces the fuzzy
grassroots ontology, and last but not least, Subsection 3.4 highlights the dashboard with its Topic Maps.
3.1. Architecture
In [34], Portmann outlined the key requirements for online reputation analysis as follows: react to mentions, put
them in context, and edit them. The present subsection presents a conceptual model that defines the structure and
behavior of this framework. Its architecture permits online searches to find data on an organizations’ reputation.
Using this framework, it is possible to scan the Web according to a query, in order to determine topic classes along
with their related tags and, thus, to identify hidden information.
Fig. 4 illustrates the FORA framework’s modular architecture. This architecture consists of three key layers:
• Reputation Search Engine Layer: Based on Web agents, this layer is designed to search for Web data. It consists
of two components: first, an ontology compiler that algorithmically collects tags and converts them into a fuzzy
ontology [30]. A fuzzy ontology formally represents knowledge as set of fuzzy concepts and the fuzzy relation-
ships between those concepts. Second, a metasearch engine that sends a search query to several subjacent search
engines (i.e., not part of the framework) and aggregates their search results (i.e., hits) via their Web API into a
hit list. To this end, it splits the resulting hits into context-based hit lists or, in other words, distributes them into
context dimensions. In the course of this, it mediates between the query engine and the various external search
engines and, in the process, creates also a ranking of the hits.
• Knowledge Base Layer: In philosophy, ontologies deal with the study of the nature of being, existence or reality,
its basic categories and their relations. In computer sciences, however, an ontology is a concept that specifies the
world in terms of a set of types, relationships and properties [14]. Now, if instances are linked to the ontology, then
we speak of a knowledge base [22]. Primarily used on the Semantic Web, a knowledge base contains a collection
of instances of concepts defined by the ontology that, in turn, specifies the structure of the knowledge stored.
Our knowledge base contains a bottom-up emerging ontology (i.e., fuzzy grassroots ontology) stored as Semantic
Web triples (i.e., subject–predicate–object tuples), and a query engine to search the knowledge base and its related
instances. A fuzzy grassroots ontology is a particular fuzzy ontology driven by the recursive social process where
prosumers jointly create and manage tags to annotate Web data such as for example by collaboratively tagging.
This kind of ontologies can be seen as a first step toward the realization of the claim to process linguistic impre-
cision not only in theory [43]. Upon a closer look, Fig. 4 reveals that the fuzzy grassroots ontology is in this layer
surrounded entirely by the query engine. This query engine is the link between the reputation search engine layer
and the dashboard layer.
• Dashboard Layer: As an interactive human-centered user interface [31], the dashboard hosts the knowledge repre-
sentation and the context-based hit list. Technically they are linked by the query engine, which translates queries
into SPARQL (i.e., a Semantic Web query language). Thereby the query engine rummages around the fuzzy
grassroots ontology, and along with it, conveys the ontology-extended search query to the metasearch engine. The
extracted knowledge is visualized on the dashboard as interactive knowledge representation [46]. Simultaneously,
the query engine delivers the extended search to the metasearch engine, which, in turn, charges various search
engines with the retrieval. At this stage, the metasearch engine splits the search engines’ hits into several (fuzzy)
context dimensions and redelivers them to the query engine again. Through the context-based hit list, the query
engine displays the resulting hits. Practicable approaches for a metasearch engine in the Social Web are Web
APIs, which allow an easy and straightforward access to underlying social media.
In the context of a closed online reputation management strategy [41], notifications can be created that respond
to certain reputation-related events and the found issues can be stored. By this means it is possible to feed the or-
ganization’s reporting system (e.g., through linking to its data warehouse). However, to close the online reputation
management loop, the ascertained reputation issues must be addressed by conveying an appropriate message to the
online buzz in order to maintain a positive online sentiment [4,23]. Eventually, the issue-related social Web responses
must be controlled.
In the following subsections, all layers and associated components are explained in more detail; to this, the next
subsection introduces the reputation search engine. Note that components-related comparisons can be found in Ap-
pendix A.
3.2. Reputation search engine
The reputation search engine layer in general rests upon the principles of Web search engines. Using Web agents
they collect specific Web data (e.g., Web links, text fragments, and metadata) on the recurrently revisited websites.
The data of each site is analyzed [20], and the results of the analysis are stored and indexed for rapid searching
afterwards [2]. The meta reputation search engine is a special case of a Web search engine that sends a given query via
Web API to several social media search engines, collect their replies and pool them in different dimensions. Based on
the characteristics of social media (e.g., blogs, microblogs, social networks), it is allocated to Dey et al.’s four minimal
context dimensions identity, location, status, and time [34]. The engine’s search corpus is thereby dynamically built,
and so the relevant Web data can be spotted at run time.
However, the ontology compiler is built on the concept of Web agents that constantly crawl the Social Web, looking
for tags. These agents identify all tagged sources in folksonomies (that results from collaborative tagging) and subjoin
them into a list. Thereby the tags and its sources are also read from the corresponding folksonomies’ Web API. Since
there is no editorial supervision in collaborative tagging, typing errors can occur. If the folksonomies would simply
be used as initial data, this would lead to overlapping but barely relating terms in the underlying ontology. To avoid
this, the collected tags must be normalized. This is illustrated in Fig. 5 as a first step towards the fuzzy grassroots
ontology.
Tag normalization is a process by which the dataset X of all n collected tags are transformed to make them con-
sistent in a way they were not before [20]. This normalization (i.e., transformation of characters and use of phonetic
algorithm) is performed before the tags are further examined. Given that simpler methods should be preceded un-
til simplicity can be traded for greater exploratory power [3], in this study we used the Metaphone algorithm (see
Fig. 5. Ontology-compilation process.
Algorithm 1 Metaphone Algorithm.

1. Drop duplicate adjacent letters other than ‘c’.
2. If the word begins with ‘kn’, ‘gn’, ‘pn’, ‘ae’, or ‘wr’, then drop its first letter.
3. Drop ‘b’ if it is after ‘m’ and if it is at the end of the tag.
4. Transform ‘c’:
5. If it is followed by ‘ia’ or ‘h’ to ‘x’ (unless ‘h’, it is part of ‘sch’, in which case it is transformed to ‘k’).
6. If followed by ‘i’, ‘e’, or ‘y’ to ‘s’.
7. Else to ‘k’.
8. Transform ‘d’:
9. If it is followed by ‘ge’, ‘gy’, or ‘gi’ to ‘j’.
10. Else to ‘t’.
11. Drop ‘g’:
12. If it is followed by ‘h’ (unless ‘h’ is at the end or before a vowel).
13. If it is followed by ‘n’ or ‘ned’ and is at the end.
14. Transform ‘g’:
15. If it is before ‘i’, ‘e’, or ‘y’, and it is not in ‘gg’ to ‘j’.
16. Else to ‘k’.
17. Drop ‘h’ if it is after and not before a vowel.
18. Transform ‘ck’ to ‘k’.
19. Transform ‘ph’ to ‘f’.
20. Transform ‘q’ to ‘k’.
21. Transform ‘s’:
22. If it is followed by ‘h’, ‘io’, or ‘ia’ to ‘x’.
23. Transform ‘t’:
24. If it is followed by ‘ia’ or ‘io’ to ‘x’.
25. If it is followed by ‘h’ to 0.
26. If it is followed by ‘ch’ drop ‘t’.
27. Transform ‘v’ to ‘f’.
28. Transform ‘wh’:
29. If it is at the beginning to ‘w’.
30. If it is not followed by a vowel then drop ‘w’.
31. Transform ‘x’:
32. If it is at the beginning to ‘s’.
33. Else to ‘ks’.
34. Drop ‘y’ if it is not followed by a vowel.
35. Transform ‘z’ to ‘s’.
36. Drop all vowels unless it is the beginning.
Algorithm 1). Note, however, that this algorithm (i.e., simplest available method at this time) is probably not the most
accurate one.
Using a lexical database that groups words into synsets [27] helps identifying synonyms in the collected tag set.
In the course of tag normalization, two linguistic issues are resolved: homographs are detected by comparing all tags
character strings, and homophones are recognized using the phonetic algorithm. To determine phonetic similarity as
Algorithm 2 Plotting points.
1. Create a tag list from a number of seeds with a predefined depth and select one source tag.
2. Select each tag in the list except the selected tag.
3. Calculate the plotted tags that are within a given distance d to the selected tag.
4. Check the number of plotted tags that have a relationship with the current tag:
a. If no plotted tags are detected, then draw the current tag with a random position.
b. If there is one plotted tag detected, then draw the current tag with the same y but with a x value that is calculated to fit the distance.
c. If there are two plotted tags detected, then draw the current tag as one of the two intersection points of two circles whose centroids and
radii are the two plotted tags and their distances to the current tag, respectively.
d. If there are three plotted tags detected, then draw the current tag as the intersection of the three circles whose centroids and radii are the
three plotted tags and their distances to the current tag, respectively.
5. Return to step 2 for the next point.
another linguistic issue, tags are reduced to a code that is able to conform to similar tags. Since synonyms show a high
similarity, they are detected during the tagspace creation.
The second step to compile the fuzzy grassroots ontology is concerned with the creation and plotting of a tagspace.
To do that, the normalized tags are affiliated to each other, based on the collected tags similarity. Similarity is a concept
whereby a set of terms within term lists are assigned a metric based on the likeness of their semantic content. Among
others, this can be achieved by defining a similarity or by using a predefined metric that states a distance between
terms. This virtually can be created using locality sensitive hashing – where the tags are hashed so that similar tags
are mapped to the same set with a high probability – and collaborative filtering – where several users define tags and
their relations jointly [34]. Using statistical means such as co-occurrences is another possibility.
For a classification of data with qualitative characteristics (i.e., nominal data), coefficients such as for example the
Jaccard, Dice, Kulczynski, and Tanimoto coefficients, are widely used [2]. The simplest metric-based coefficient that
can be used to measure semantic correlation between tags is according to Hassan-Montero and Herrero-Solana the
Jaccard similarity coefficient [17]: Let A and B be the sets of resources characterized by two tags, then
|A ∩ B|
sJ (A, B) =
|A ∪ B|
where | . | denotes the number of elements in a given set. In other words, relative co-occurrence is identical to the
partition among the amount of resources in which tags co-occur and the amount of resources in which either of the
two tags appear. This collection method causes tags to become united and offers a semantically consistent picture
(i.e., tagspace), in which, nearly all of the tags are related to each other. Even in its simple form, sJ allows identifying
synonyms, as introduced in the previous subsection. In the course of this, finally, the synonyms can be matched with a
lexical database (e.g., WordNet [27]), thereby allowing a more accurate or fine-grained determination of synonymity.
Note that from the Jaccard similarity sJ , its distance can be calculated by dJ = 1 − sJ = |A ∪ B| − |A ∩ B|/|A ∪ B|.
However, other distances d, based on other norms, can also be used. For a comparison of distances see Appendix A.1.
To begin plotting the point representation of the tagspace, it is necessary to set a limitation for the tagspace.
Algorithm 2 starts with a number of seed points and the child point locations are computed based on Bourke’s classical
algorithm, which calculates the intersection of two or three circles [7].
After the normalized tags have been combined, assorted and plotted onto a tagspace, a computer-understandable
ontology can be established. Algorithm 2 allocates the position of each point in the tagspace. Based on this algorithm,
the necessary points in the selected region can easily be shown. Incidentally, this is, among other things, very effective
for supporting a zoom function [34]. Another parameter to take into account is the constant variability of the under-
lying data; often data are static, but here, they are constantly moving around. In fact, they change every week, hour
or second (depending on the Web agents update frequency). This consideration is legitimate because social Web data
come from real world, where no absolutes exist. Online trends can change extremely rapidly. To interact with live
data, the tagspace needs to be continually updated. As a result, the introduced approximate plotting algorithm is able
to provide a good perspective on moving data [23,24].
The ontology adaption is the last step of the fuzzy grassroots ontology-compilation process that separates the
plotted tagspace into different clusters with the help of a fuzzy clustering algorithm [1,31]. In order to detect a re-
spective algorithm for the FORA framework, a comparison of three different algorithms is drawn: the FCM, the
Gustafson–Kessel (GK), and the FLAME algorithm. To evaluate the most useful one, suitable factors are needed.
In order to find such factors, all three algorithms are examined for potential comparable properties. Since the algo-
rithms are mathematically comparable, the factors must be measurable and verifiable. This is guaranteed twofold:
First through mathematical comparisons, and second by Matlab tests. Thus the factors for the comparison are com-
plexity, permanence, and adaptability. Appendix A.2 illustrates the selected factors (with dedicated weight and value
range), and shows the reasons FLAME come out victorious. Appendices B.1 and B.2 additionally show our adaption
of FLAME in Matlab code. More information concerning this comparison (i.e. point of departure, evaluation, and
results) may be obtained from Portmann [34].
For the ontology adaption all n tags plotted on the tagspace will be sorted by the Fuzzy clustering by Local
Approximation of MEmberships (FLAME) algorithm [12]. This algorithm is based on an approximation of neighbor
tags in the tagspace. FLAME starts with the identification of cluster supporting objects that are representative tags
around which the cluster will be constructed. To do this, the similarities between each pair of tags and the k-Nearest
Neighbor (kNN) are picked to weigh the density
1 1
ρ= = ∀x
d n
j =1 aij (xi − xj )
2
around the representative tags. For this purpose each tag is connected with its kNN, with k being a predefined constant:

ai,j = 1, if xj ∈ xi
∀i, j
ai,j = 0, else
The sums of the distances d of these connections are used to calculate the density ρ at each point. Any point whose ρ
is greater than that of all its neighbors is called cluster supporting object CCSO . Outliers are likewise determined; they
possess a smaller ρ than a predefined threshold ε. Fu and Medico [12] perform a comparison on how the algorithm
reacts in partitioning of genes. We applied their finding to our ontology adaption. To this end, in Appendices B.1 and
B.2 our used FLAME algorithm calculations (i.e., explaining k and ε) may be found. Anyway, now we have c − 1
clusters ΓCSO and an outlier cluster ΓO . The outliers are given full membership to ΓO
∀i ∈ ΓO : uc+1 (xi ) = 1
∀i, l ≤ c : ul (xi ) = 0
and the cluster supporting objects full membership to their own clusters ΓCSO
∀i ∈ ΓCSO : uli (xi ) = 1
∀i = j : ulj (xi ) = 0
where li is the cluster with i as the CSO. Each remaining point is now allocated to all clusters Γi with the same
membership grade
1
∀x ∈
/ (ΓCSO ∨ ΓO ), ∀l(x) =
c+1
Note that the sum of all membership grades of a point is exactly one
c

∀i ul (xi ) = 1
l=1
and no cluster is empty
n

∀l ul (xi ) > 0
i=1
For the definition of the weights, the tags xj that have higher similarity must feature higher weight factors. So in
the algorithm the weights
Algorithm 3 FLAME Algorithm.
1. Extract the tagspace-inherent structure by constructing a neighborhood graph where each tag xj is connected to its kNNs.
2. Calculate a density ρ for each tag based on its distance d to its nearest neighbor using the density formula ρ = 1/d.
3. Assign each tag:
a. If the tag has a higher density ρ than all its neighbor tags to a cluster supporting object CCSO = {C1 , . . . , Cc−1 }.
a. If the tag has a density lower than all its neighbors, and lower than a threshold ε (i.e., an arbitrary constant) to cluster outliers CO (= Cc ).
4. Assign the tags of the cluster supporting objects CCSO full membership to its respective clusters ΓCSO = {Γ1 , . . . , Γc−1 }.
5. Assign the tags of the cluster outliers CO full membership to its respective cluster ΓO .
6. Assign for every remaining tag membership to all clusters Γi (incl. the outlier ΓO ).
7. Reiterate assigning membership until the approximation error E({uij }) converges to zero by updating the memberships uij of the remaining
tags by a linear combination of the memberships of its nearest neighbors (assign higher weights wij for closer objects xj ).
sij
wij =
xm ∈ kNN(x i )sim
define to which extend each neighbor contributes to approximation of the membership of that neighbor with wij being
the weight of j on i; the sum of all kNN weightings is 1. Regardless of the underlying metric, sij is a non-normalized
similarity. In the iterative part of Algorithm 2 we calculate the corresponding coefficients

∀i, l : ul (xi ) = wij × ul (xj )
xj ∈kNN(xi)
In the course of this process, we arrive at an approximation error,
2
E {uij } = u(xi ) − wij u(xj )
xi ∈X xj ∈kNN(xi )
whereby u(xi ) represents a membership grade for each CSO: ul (xi ) for l = 1, . . . , c + 1.
Fu and Medico demonstrate that the algorithm converges in the first place, in a heuristic way [12]. So, after the
algorithm converges, the ontology can be processed. There are three different methods to process the ontology: first,
a point belongs to the cluster to which it has the highest membership (i.e., yielding a Boolean clustering); second,
a point belongs to all clusters it has membership higher than a predefined threshold (i.e., yielding a fuzzy clustering);
or third (also yielding a fuzzy clustering), the data are processed as a whole (i.e., any membership of any data point to
a cluster is stored as ontology together with the corresponding coefficients). The FORA framework supports this last
approach; all data points and clusters are stored using a knowledge administration system (i.e., a graph database).
Hence, the fuzzy grassroots ontology contains the clustered tagspace from which hyponym–hypernym knowledge
structures can be derived, with a membership degree of each tag indicating the part-of-strength of the tag to a certain
class. Again, following Baker [3], by assuming that simple methods should be preceded until simplicity can be traded
for better reason, the class is named by the tag closest to the centroid of the class [34]. Using inductive fuzzy clas-
sification, in future, however, we try to advance the naming of the respective class (see [23,24,35]). All the tags thus
form the ontology and the named (as mentioned above) resulting fuzzy clusters that are translated to an OWL 2 syntax
(i.e., a recent Semantic Web language), as Bobillo and Straccia [6] propose. The created fuzzy grassroots ontology
consists of a quintuple F = X, C, T , N, A [30] where X denotes the set of all n normalized tags; C the set of all
fuzzy clusters; T the fuzzy taxonomy relations among the set of clusters C; N the set of non-taxonomy fuzzy asso-
ciative relationships; and A the set of axioms expressed in a proper logical language (i.e., predicates that constrain the
meaning of concepts, objects, relationships and functions). Each cluster C is a fuzzy set on the tagspace, and the set
of units U of the fuzzy ontology is defined by U = C ∪ X. The fuzzy taxonomy relations T organize the concepts into
sub-and-super-concept hierarchic structures, and through naming each resulting relationship the non-taxonomy fuzzy
associative relationships N relates the units U across the hierarchies. In the course of this, the relationship T (i, j )
indicates that a child j is a conceptual specification of a parent i with a fuzzy membership degree uij (see Subsec-
tion 2.3). N denotes the non-taxonomic fuzzy associative relationships that relate entities across the tree structures
(e.g., naming, locating and functional relationships). Naming relationships connotes describing the names, locating
Fig. 6. An exemplary fuzzy ontology schema.
the relative location, and functions (or properties) of concepts (i.e., clusters C). Adapted from Parry, Fig. 6 should be
exemplary for the underlying but rudimentary graph database schema of the fuzzy grassroots ontology [30].
Relational databases with their underlying schemas make some queries very easy and others more difficult (see
[38]). However, in todays Web, we want to connect data as the domain stipulates, thereby allowing structures and
schema to emerge together with our growing understanding of the environment, rather than being forced upfront,
when we know least about the real shape and intricacies of the underlying Web data.
Note, however, that in the current fuzzy grassroots ontology, until now, based on similarity only part-of-
relationships (e.g., isA) can be discovered. Using inductive fuzzy classification [23,24,35], however, we seek to amend
the ontology in the near future. Following Bobillo and Straccia’s approach, a procedure to represent the ontology-
inherent information within the current standard languages and tools is put forth [6]. Eventually, the resulting fuzzy
grassroots ontology now only needs to be stored using Semantic Web triples, with a class Tag, an object property
isPartOf/isA and an annotation property distance with a [0,1] range, using the OWL 2 syntax.
The FORA frameworks metasearch engine is a search engine that sends the fuzzy grassroots ontology-enriched
query to several subjacent search engines and aggregates the results into lists of context. The enrichment of the user
query is based on the knowledge base administrated through the fuzzy grassroots ontology. The fuzzy grassroots
ontology is queried using SPARQL. This query language allows returning all terms that are at a distance closer than a
weight K. In Subsection 3.4, we illustrate how the user graphically specifies K using SARQLs FILTER function. The
result would be to retrieve every tuple of the knowledge base that participates in the assertions to a special degree (e.g.,
K ≈ 0.3 → ?distance < 0.3). Next, via their Web API, the metasearch engine takes the enriched query, passes
it to several other heterogeneous search engines and then compiles their results in a homogeneous manner. To this
end, the metasearch engine decomposes the enriched query into subqueries for submission to the constituent search
engines. Because various search engines employ different query languages, the metasearch engine applies wrappers
to the subqueries to translate them into the appropriate query languages of the underlying search engine [2,36]. Last,
the metasearch engine collects their replies and classes them with context dimensions.
Now, to use the generated fuzzy grassroots ontology, it needs to be stored in an appropriate knowledge administra-
tion system. The next subsection is therefore concerned with the knowledge administration of the FORA framework.
Fig. 7. The reputation search engine.
3.3. Fuzzy grassroots ontology
The knowledge base layer is a collection of instances of the concepts defined in the fuzzy grassroots ontology, and
the fuzzy grassroots ontology itself specifies the structure of the knowledge stored in the knowledge base. Together, the
instances and the fuzzy grassroots ontology make up the knowledge base, which is the FORA framework’s core [22,
23]. Since the framework is based on the reputation search engine, the knowledge base represents the fuzzy grassroots
ontology plus ad hoc Web content.
The last subcomponent of the ontology compiler separates the plotted tagspace into clusters with the help of the
fuzzy clustering algorithm FLAME, spawning the fuzzy grassroots ontology. Using Semantic Web technologies, this
ontology is stored in a knowledge administration system. Based on a high-performance and persistent graph database,
the underlying store provides a solid storage layer that enables to scale to billions of triples. For a comparison of
knowledge administration system see Appendix A.3. In fact, the store is designed to manipulate triples that can be
queried with SPARQL (see Subsections 3.2 and 3.4). By embedding related tags, whose quality is user-defined by
means of positioning a slider that triggers a SPARQL FILTER function, the original query is enriched.
Applied to the rudimentary graph database schema, a respective FILTER query could be as
SELECT ?apple WHERE

{ ?apple ex:isA fruit
?apple ex:distance ?distance
FILTER (?distance<0.3)}
returning, by way of example (of Subsection 3.2), Granny Smith.

The query engine is a crucial point of the FORA framework and acts as a hub between the fuzzy grassroots
ontology, the metasearch engine and the dashboard. It translates user queries into SPARQL queries and sifts through
the fuzzy grassroots ontology. Following this, it sends the enhanced query to the metasearch engine that decomposes
the query into subqueries for submission to the underlying search engines. This is illustrated in Fig. 7. Collected by
the metasearch engine, the query engine finally provides the dashboard with the classified results; this is presented in
the following subsection.
3.4. Dashboard with Topic Maps
Based on the previous analysis, an intuitive dashboard was created that allows communication operatives to quickly
find salient documents. The query engine conveys the context classified to the dashboard, which performs online
reputation analysis within the scope of the FORA framework application. In response to a users interaction with
the dashboard, it changes the provided content. Furthermore the dashboard provides various functions such as stor-
ing, triggering and reporting search results. Nevertheless, its key elements are the knowledge representation and the
context-based hit list.
Visualization techniques empower users to spot patterns in social Web data, identify therein areas that need
additional analysis and make sophisticated decisions based on these patterns. The human capability to converse,
communicate, reason and make rational decisions in an environment of imprecision, uncertainty, incomplete informa-
tion and partial truth will be supported by this visualization. An interesting feature of this visualization is the ability
to discover hotspots through a possibility of interaction [20,31]. To increase the ability to explore data (and thus, to
better understand results), an effective integration of the visualization and interaction applications is important. The
field of analyzing data to identify relevant concepts, relations, and assumptions, combined with the conversion of data
into machine language, is known as knowledge representation [46]. Its aim is to present data in a manner that will
enable reasoning. On one hand, the fuzzy grassroots ontology provides FORA with a general knowledge of vague hu-
man concepts and, on the other hand, the fuzzy grassroots ontology-based interactive visualization of this knowledge
through a visualization technique helps users to identify patterns [1]. Thereby Topic Maps [32] are selected as good
solutions to visualize knowledge (see Section 1). In Appendix A.4 a comparison between knowledge representation
systems is made.
Today, browsing and searching are the main interactions on the Web. While browsers provide visual mechanisms
for navigating the Web, in many cases, search engines are the place where a Web navigation process starts. The
FORA framework, however, provide possibilities for searching and browsing not only the retrieved Web documents,
but also their underlying structural semantics [35]. Consider as example a user searching the Web. Now, in addition to
browsing only from a found Web document via a link to another document, his search is extended by the possibility
to follow a link into visualized structural semantics. The user navigates the knowledge structures for some period of
time and then returns to a link to a different Web document [5]. This is the FORA framework’s underlying principle.
In philosophy we may distinguish between perceptions and conceptions, the former is what humans can perceive,
the latter the generalization of the perceived. Only generalizations can be modeled as ontologies. Because social Web
users apply natural language to describe their perceptions, a perception-based ontology can be created. A Topic Map is
an ontology visualization, which primarily consists of a set of topics and associations. The topics thereby are modeled
statements of real world subjects, and the associations’ general relationships between the modeled statements [32]. In
the proposed ontology adaption, a cluster center Γi pertains to a topic and a fuzzy partition matrix U = [uij ], with
i = 1, . . . , n and j = 1, . . . , c, what is equal to μΓ i (xj ) ∈ [0, 1]. Each element uij tells the degree to which element
xj belongs to cluster Γi , and it defines the associations with further topics (and associated terms).
Studies reveal that the attention allocated to online content is strongly biased [47], so that while most Web doc-
uments get a small attention span, only a few grow really popular. This implies that potentially relevant information
might be missed. For that very reason, viable solutions must be searched to help communication operatives “allo-
cate their attention effectively among the overabundance of information sources that might consume it” [40]. This is
known as economics of attention, and leads directly to how users access key information in information overloaded
Web [19]. On the one hand this addresses how to browse and find the most prominent documents in big corpora,
and, on the other hand, how to quickly discover critical knowledge from these documents and situate them within the
big picture. Using Zadeh’s z-mouse concept [50], with a zooming in-and-out function (see Subsection 3.2) thereby
more or less associated topics can be visualized; in other words, by varying a search weight K via a slider (see Fig. 9)
a communication operative might interactively control the inclusion of visualized topics to be presented on the dash-
boards knowledge representation [34]. Topics farther away visualized from other topics are thus less related and also,
terms visualized farther away can be considered as less relevant (see Section 1). This is a special form of interactive
data mining [20].
The appearance of the dashboard includes two parts: first, on the left side, the user-navigable visualization of the
grassroots ontology as Topic Maps, and second, on the right side, the context-based hit list found by the search engine
(see Fig. 8). The context-based hit list thereby constitutes the aggregation part of the framework, where the underlying
instances of the Topic Map visualized fuzzy grassroots ontology is presented to the user.
Dey et al. introduced four essential characteristics of context information – identity (i.e., who?), location (i.e.,
where?), time (i.e., when?), and status (i.e., what?) [9]: Identity refers to the ability to assign a unique identifier; lo-
cation deals with information that can be used to deduce spatial relationships; status identifies intrinsic characteristics
Fig. 8. The FORA framework dashboard.
that can be perceived; and finally, time helps temporally characterize a situation. In the Social Web the underlying data
change constantly, which is why no simple hard but rather a fuzzy allocation to context dimensions can be made. So,
for example, the identity dimension can consist of fuzzy stakeholder groups; location of fuzzy geographical assign-
ments; status of fuzzy annotations; and finally time of fuzzy periods. Through a splitting of aggregated information to
these context dimensions, a user can easier understand and evaluate the information. Up until now, the different kind
of social media serves as discrimination (see Section 4). Note, however, that at the moment we are focusing to shift
this limitation.
So what really makes the dashboard unique is what happens after a communication operative enters his search
query. Instead of delivering millions of hits as a long list, the context-based hit list arranges similar ones together
into context-dimensions and the knowledge representation appendant represents interactively the fuzzy grassroots
ontology-inherent knowledge. The contextual cues actively support this phase that can help communication operatives
encounter information through browsing (see [16]). In that way, it is possible to search results by topic and thus a
communication operative can zoom in on exactly what he is looking for or discover unexpected relationships between
hits (see hinted zoom function in Subsection 3.2). The individual sources are accessed via Web API. Following Zadeh’s
z-mouse concept [50], this zoom function is what reflects underlying fuzziness visually. Based on this zoom function,
the SPARQL FILTER function picks more or less corresponding tags (i.e., query enrichment; see Subsection 3.3).
Rather than scrolling through page after page, the dashboard helps find results that may have been missed or that were
buried deep in a ranked list. Thus, using the FORA framework, it is possible to semantically analyze the Web, and
determine query-related topic classes with related terms and thus potentially hidden reputation issues.
In the next section, the YouReputation prototype will be introduced that represents an implementation of the pre-
sented FORA framework. Finally lessons learned that have been gained during the implementation process will be
given.
4. YouReputation prototype and lessons learned
By posting or commenting Web data, the Social Web made it easy to malign an organization. Scanning and moni-
toring helps to quickly track these kinds of data and, maybe after a forecasting of an assault’s impact, an organization
can respond in a timely manner and hence better manage its reputation. By this means, the distance between an orga-
nization and its customers can be minimized. Now the YouReputation prototype gives organizations a hand to do an
effective online reputation analysis. The implemented prototype thereby acts as proof of concept of the FORA frame-
work and is in a continuous improvement phase. To support the framework, also several other prototypes are built as
proof of concept of subparts of the framework (e.g., see [24,35]). For the creation of the prototype’s ontology, the lex
parsimoniae is followed. This law of parsimony suggests tending towards simpler solutions until some simplicity can
be traded for increased explanatory power [3]. This goes hand in hand with the applied methodology of (rapid) pro-
totyping, which entails compromises in functionality and performance in exchange for enabling faster development
and facilitating application maintenance [33]. The YouReputation prototype is provided as a Web service available
Fig. 9. The YouReputation dashboard (from www.youreputation.org).
at www.youreputation.org. The prototype automates the process of online reputation analysis by tapping prosumers’
online buzz and distilling what they are talking about.
In the following subsection, we first introduce the YouReputation prototype, in Subsection 4.2, we present a use
case, which emphasizes the prototypes advantages, and last but not least, in Subsection 4.3, we evaluate the prototype.
4.1. The YouReputation prototype
The YouReputation prototype underlying fuzzy grassroots ontology is created with Matlab to allow matrix manip-
ulations (i.e., normalized input from folksonomies; see Subsection 3.2), plotting of data (i.e., compiling a distance
matrix), implementation of algorithms (i.e., fuzzy clustering algorithm), as well as interfacing with third party pro-
grams (i.e., AllegroGraph’s RDFStore [11]).
Fig. 9 illustrates the prototypes dashboard for a search for Apple Inc. For clarification, this figure merely shows
the mentioned concept of browsing semantic knowledge structures [24,35]. Note that with the current algorithm of
the YouReputation prototype, experts performing online reputation analysis, possible less-interesting topics as OSX
or Mac may be found, which should be filtered in the future.
However, instead of endless and often confusing hit lists, the YouReputation’s dashboard provides relevant con-
tributions in clear manner and thereby meets the organizations first requirement (see Subsection 3.1). The dashboard
thereby is HTML-based, and is rendered in a Web browser. The YouReputation prototype recognizes the correlations
of the keywords by its fuzzy grassroots ontology, and arranges similar results to bundles. This fulfills the second re-
quirement. So far, following a design-oriented approach, we implemented the context dimension time (i.e., when?),
and status (i.e., what?). Following the law of parsimony, to explain and illustrate, we used the Web APIs of Delicious
(i.e., status dimension) and Twitter (i.e., time dimension) [34]. In future, we plan to improve and automate this pro-
cess and thereby include additional context dimensions. If there is a demand to respond immediately to issues, the
prototypes alert functionalities, regular summaries and a simple forwarding are very important. These functionalities
can be addressed with corresponding Web services, by which the third and last key requirement is met [34].
In a use case, in the next subsection, we present the advantages of the YouReputation prototype.
4.2. Apple Inc. use case
To analyze their reputation, Apple Inc.’s communication operatives want to know all mentions about their brands,
services, products and executives. Hence, to spot early dissonance, a sound online reputation analysis should include
fuzzy indications not directly mentioned by the underlying social media. In addition, during the analysis, issues should
be summarized into related topics. The problem with new and previously unobserved information on the Web is that
the relationship between terms and topics is not precisely known (see Subsection 2.3).
Starting in June 2012, the tag “BoycottApple” became very popular on Twitter. This tag was about Apple’s in-
junction against the Galaxy Nexus flagship Android smartphone. Although an appellate court has lifted the injunction
end of July 2012, the injunction made BoycottApple a long trending tag in social media, which still lasts. Since it is
common practice on social media services to use more than one tag (e.g., BoycottApple) with additional tags (e.g.,
Samsung, GalaxyNexus, Android, Hypocrite, etc.) it is possible to put these tags (i.e., terms) in relation to each other
and, hence, cluster these tags to topics related to each other (e.g., in a tagspace; see Subsection 3.2) – and that from
the point of first common mention of the tags. So, if Apple reacts at a rapid pace, it could turn away a crisis before it
really becomes a no-brainer eventually making the headlines. As we have seen in Subsection 2.3, in the beginning an
organizations sphere of action is high, dealing with a certain issue is not too costly, and the public attention is rather
limited. But the more the public awareness increases, the less the organization can do in order to solve the problem
the easy way.
Now, to scan mentions in relation to the organization, a communication operative first enters the search term Apple
in the search field box on the start page of the prototype. On the left side of the dashboard, immediately, a fuzzy
grassroots ontology-based knowledge representation (i.e., Topic Maps visualizing all Apple-related topics and terms;
see Fig. 9) appears.
The more two or more tags were used together, the closer they are mapped to each other in the underlying fuzzy
grassroots ontology. The frequency of co-occurrence thus determines the fuzzy membership degree – from 0 if they
never appear together to 1 if they always appear together (see Subsection 3.2 et seqq.). The fuzzy grassroots ontology
constitutes the attempt to capture the vagueness of human concepts. The visualized return, as Topic Maps should be
considered as visualization of the fuzzy grassroots ontology. Through the dashboard’s zooming-in-and-out-function
(i.e., moving along certain term’s fuzzy membership line), the Topic Map allows to observe more or less related terms.
So, for example, first only Apple, Samsung and Android are visualized as Topic Map. By zooming-out of the topic,
Galaxy Nexus and Hypocrite may arrive on the Topic Map dashboard. In the curse of time, the fuzzy memberships
(i.e., frequency of co-occurrence) can move gradually, which can be observed on the dashboard’s Topic Map. After
scanning online buzz, in the monitoring phase, this may be an indicator – a topic improves or worsens, and could be
taken as a starting point for an adequate reaction by conveying an online message (see Subsection 2.2).
On the right side of the dashboard, a hit list that is partitioned into context dimensions appears. These context
dimensions come across as any information (no matter how fuzzy) that characterizes a situation related to the inter-
action of communication operatives with the prototype and the underlying social media. In this way it is possible to
browse the ontology and simultaneously discover underlying social media sorted according the context dimensions.
To illustrate, the hits for the status dimension are coming from Delicious service because Delicious includes mostly
descriptive information. However, for the time dimension, the hits are coming from Twitter service because Twitter
provides additional data such as the date and time of a tweet. Thus, it is possible to find past as well as real-time
information. Thereby these two exemplary social media are connected to the prototype via their Web APIs.
The advantage is that communication operatives not only find more relevant information concerning their entered
search term but also receive more structured information (based on the ontologies and the context dimension parti-
tions). Their application queries several top social media search engines, combines the results and generates a hit list.
With a conventional Boolean search system, they would only find information containing the term Apple. In contrast,
the prototype enables to find not only the search term but also more or less related topics and terms. Since the dash-
board is interactive, Apple’s communication operatives can zoom-in-and-out and with the help of Topic Maps, browse
the ontology in a straightforward manner.
In the next subsection the YouReputation prototype’s evaluation is presented in a nutshell.
4.3. Evaluation
The YouReputation prototype provides proof of the FORA framework. However, here, the YouReputation proto-
type itself is evaluated. Since the fuzzy grassroots ontology depends on clustering techniques, it is obvious to apply
its validation techniques (i.e., analyzing with internal, external, and relative criterion; see [34]). There are no bench-
mark data to compare the FLAME algorithm obtained cluster quality, though, from a human perspective (i.e., expert
aspect), the quality of the clustering results looks well (i.e., relative criterion). Yet, since in future the fuzzy grassroots
ontology gets updated automatically, it can barely be benchmarked automatically, at this point an external criterion is
abstained from.
Objective functions in clustering formalize the purpose of attaining high intra-cluster similarity and low inter-
cluster similarity (i.e., internal criterion). For a good clustering (i.e., with high information entropy) the David–Bouldin
index should be as low as possible. For our crawled data the index is below 1; thus, this outcome also is considered
as quite well [34]. However, a good outcome on an internal criterion does not automatically translate into good ef-
fectiveness in an application. An option to internal criteria is the direct evaluation of the application of interest. This
relative criterion is examined with the evaluation of the YouReputation.
Together with communication operatives, we qualitatively evaluated the prototype using a discount usability
method [28]. To do this, the prototype’s functionalities were evaluated in terms of how well they measure up against
comparable applications (i.e., Actionly, MeMo, and Sysomos) on the market to the communication operatives’ re-
quirements (see Section 2).
In contrast to the YouReputation prototype, Sysomos, MeMo news and Actionly provide all broad implemented
functions. In the following we present a shortened evaluation of the communication operatives of PostFinance (i.e.,
a Swiss financial service unit of Swiss Post). Moreover, Appendix A.5 highlights the comparison of the prototype
with these closest competitive systems available on the market. (Please see Portmann for full length evaluation [34].)
Instead of endless and often confusing hit lists the YouReputation’s dashboard applet provides relevant contri-
butions in clear manner and thereby meets the requirements of communication operatives for reasonable manual
expenditure and final qualitative reports. In contrast to other applications, the search for opinion leaders, new and
relevant topics and important keywords is no longer necessary. Yet, the YouReputation prototype recognizes the
correlations of the keyword by its knowledge base, and arranges similar results to bundles. In this way, the online
reputation analysis requirement to put mentions in context is complied. Moreover, the YouReputation prototype is
ahead of others in terms of supporting semantic analysis. However, it must be said that the prototype works so far
only for certain search terms and new keywords have to be learned first. Through the implementation of a continuous
automatic update, the prototype would learn the keywords fully automated.
On the other hand, a simple user experience under load test was run. For the test the reviewers used the dashboard
from different locations for roughly the same time. Thereby each reviewer independently entered unspecified search
terms and stopped the time to load the Topic Map and the dedicated hit list. On average (i.e., out of 30 trials), it takes
1.05 seconds to load the Topic Map (shortest loading time was 0.9 seconds, the longest 1.4 seconds), and 3.86 seconds
to load the hit lists (shortest loading time was 2.2 seconds, the longest 4.8 seconds). Because the hit list widget depends
on the knowledge representation widget, it makes sense that the load of the hit list is slower. Notwithstanding there is
still development potential, but following the law of parsimony, for the YouReputation prototype these loading times
were regarded as good enough [34].
In summary, it can be stated that none of the applications on the market meet the requirements of communications
operatives in full. All conventional applications must expand their semantic functionalities towards YouReputation,
and YouReputation, in turn, its user-friendliness and reliability towards the products on the market (see [34]).
5. Conclusions and future directions
This article provided a foundation for further analysis of a reputation system. The introduced FORA framework is
an approach to gain deeper insights into an organizations online reputation. The social Web can reveal what customers
actually think about an organization, so whether they tweet, comment or blog about it, with our framework an organi-
zation will make sense of its customers’ opinions. And because the boundaries in fuzzy classification systems are not
rigorous, it is possible to find more and higher-quality results based on the presented fuzzy grassroots ontology. With
this in mind, among other things, the YouReputation prototype is intended to illustrate the possibilities provided by
the vast amount of recent social Web data. Yet, it is quite evident that the YouReputation prototype underlying fuzzy
grassroots ontology can also be used for other purposes, such as for example digital privacy protection [36] and crisis
and stakeholder management [37]. At the moment, the idea to apply the fuzzy ontology to competitive intelligence is
under way. Note that a combination of the presented ontology with the (dis)trust score of Victor et al. [44] might be a
promising research direction.
The prototype attempts to establish a knowledge representation for reputation analysis through Topic Maps –
a method for representing and exchanging knowledge. These visualized interactive Topic Maps help find similar
linguistic terms clustered around a topic. The interactive visualization of these maps is a first step towards opening
up communication operative horizons. However, in the future, through induction [23], the maps underlying semantic
fuzzy grassroots ontology ought to be enhanced to allow automatic reasoning. So, in a computational semiotic sense,
we expect to extend the fuzzy grassroots ontology to a fully automated thesaurus containing gradual synonyms and
antonyms (see [24,35]). Using the social Web user’s loose words drawn from natural language, these enhanced maps
should allow a causal reasoning that is perception-based to better mimic the human way of thinking [29]. Hence,
springing from paraconsistent logic, a future extension may rather be able to capture (and not only process) words,
perceptions, and propositions [49]. Kosko’s fuzzy cognitive maps – a method for representing loose causalities – could
effectively enhance the Topic Maps based on common sense [26,36].
To this end, in addition, loose causalities have to be collected from the Web. We think that the dashboard’s hit list
underlying context dimensions could already include such causalities. Therefore future research will include Dey et
al.’s four context dimensions [9] that not only allow to straightforwardly interact with different kinds of social media,
but also to naturally sort fuzzy graph-structures for representing causal reasoning. For each dimension, that is to
say, assorted individual fuzzy weights that present relationships between different concepts of the respective context
dimension could be deduced. At the moment, this is under evaluation.
Another important point for further improving reputation management is the analysis and inclusion of word-
inherent positive or negative connotations to automatically categorized Web sources [41]. To this end, a comprehensive
framework for social media elements is at the early stage of development. With this in mind, a linguistically enhanced
(fuzzy) partitioning of the data into positive, negative, and irrelevant examples of a rule could be used as defined by
Sudkamp in [42]. In addition, the patterns of dashboard interactions could be used as a springboard to affect the future
ranking of documents, producing a more personalized presentation of results.
In the process of online reputation management, after an analysis of its reputation, an organization should enter
an appropriate online conversation with its stakeholders. By being a part of a social Web, an organization is a part of
the conversation. So it can talk back to customers, respond to their comments and answer their questions. However,
often the organizations hang back to communicate with stakeholders by social media [13]. But to hold this media in
low esteem is to obstruct the organization’s business and if they miss the chance to reorient to their detractors finest
line of attacks it surrenders its scope. If an organization wants to safeguard its image to use a word to mean Oscar
Wilde, it needs to reconsider its online reputation management strategy and recognize that there is less control over
an organization’s messages than just a few years ago. It may have to deal with unknowns that can spatter the repu-
tation overnight. However, no organization will ever escape its online reputation; rather the repartee is permanently
ongoing.
Acknowledgement
Upon the recommendation of the referees this article is a revised version of Portmann’s article [33] presented at
the Sixth International Summer School on Aggregation Operators in Benevento, Italy. That is why we are grateful to
the referees for this honor. We also appreciate our colleagues of Mediamatics, the Fuzzy Marketing Methods, and the
Berkeley Initiative in Soft Computing research centers who selflessly helped with word and deed when we sometimes
could not see the wood through the trees. Thereby the Swiss National Science Foundation supported this research
under grant number PBFRP2-138628.
Appendix A
The following comparisons have been taken from [34]. There also more information about our approach and in
depth explanations of the results can be looked up.
A.1. Comparison of distance measurements
Distance (continuous space) Distance (discrete space)

• L1 (or City-block, Manhattan or Taxicab (for p = 1)) • Dice (or Simple Matching)
2|A∩B|
dL1 (x, y) = ni=1 |xi − yi | dS (A, B) = 1 − |A|+|B|
• L2 (or (Squared)
Euclidean (for p = 2)) • Jaccard
dL2 (x, y) = n (x − y )2
i=1 i i dJ (A, B) = 1 − |A∩B|
|A∪B|
• L∞ (or Chebyshev, Maximum (for p = ∞)) • Kulczynski
dL∞ (x, y) = maxi (|xi − yi |) dK (A, B) = 1 − 12 ( |A∩B| |A∩B|
|A| + |B| )
A.2. Comparison of fuzzy clustering algorithms
Factor Weight FCM GK FLAME

Complexity 50% 1 1 1
Permanence 30% 0 0.5 1
Adaptability 20% 1 1 0.5
Total 100% 0.70 0.85 0.9
A.3. Comparison of knowledge administration systems
Factor Weight AllegroGraph Jena KAON

Number of supported languages 20% 0.6 0.6 0.4
Speed of retrieval 25% 0.5 0.75 0.75
Type of storage 20% 1 0 0
Backup-management 20% 1 0.5 0.5
Version control 5% 0 0 0
Methodological support 5% 0 0 0
Automatic classification 5% 0 0 0
Total 100% 0.65 0.41 0.37
A.4. Comparison of knowledge representation systems
Factor Weight RDFS OWL Topic Maps Tag Clouds

Visualization 20% 0 0.75 0.5
User incentive 15% 0 1 0.5
Comprehensibility for user 15% 0 1 0.5
Standard 12% 1 0.5 0
Modeling method 9% 0 1 0.75
Critical size 9% 0 1 0.5
Extensibility 12% 1 0.5 0
Semantic expressiveness 8% 1 0.5 0
Total 100% 0.32 0.79 0.36
A.5. Comparison of selected applications
Factor Sysomos MeMo Actionly YouReputation

1. Usability
Look & feel Excellent Poorly Good Simple
Forward & share Yes No Yes API
Respond Yes No Yes Yes
Data export Spreadsheet No Spreadsheet Spreadsheet
2. Languages
Language choice DE, EN, FR, IT DE, EN, FR, IT DE, EN, FR, IT EN
3. Quantitative analysis
Graphics Excellent Good Good Good
Duplicates & spam Yes Yes No Yes
4. Qualitative data analysis
Keywords Yes No Yes Yes
Key topics Unreliable No Unreliable Yes
5. Reporting
Alert functions Email No Email API
Overview by email Daily Daily Daily No
Appendix B
The following algorithms have been taken from [34]. There also more information about our approach and in depth
explanations of the results can be looked up.
B.1. Matlab calculation of the memberships of the clusters
Function [CluMat,turn,TagList,CSOs] = flame_calc(XDat,Tags,KNN,Turns)

% FLAME_CALC calculates the memberships of the clusters
%
% written by: Sandro Kolly
% date: 2011/08/18
[~,N] = size(XDat);
[SimMat,CSOs,CluMat] = flame_init(XDat,KNN, N);
CluNum=numel(CSOs);
EyeMat=eye(CluNum+1);
turn=0;
sumep=zeros(N,1);
for epcso=1:CluNum+1
for epsim=1:N
sumep(epsim)=SimMat(:,epcso)’*CluMat(:,epcso);
end
end
% calculate the Fuzzy Memberships by Local Approximation

While turn<Turns && sum(sumep)>0.000001
turn=turn+1;
for line=1:N
if ~ismember(CluMat(line),EyeMat)
for row=1:CluNum+1
CluMat(line,row)=SimMat(:,line)’*CluMat(:,row);
end
end
end
%calculate the E(p)(LAE/NAE)
for epcso=1:CluNum+1
for epsim=1:N
sumep(epsim)=(SimMat(:,epcso)’*CluMat(:,epcso))^2;
end
end
end
TagList=cell(numel(CSOs),1);
for i=1:numel(CSOs)
TagList(i)=Tags(CSOs(i));
end
TagList(numel(CSOs)+1)= cellstr(’OutlierGroup’);
end
B.2. Matlab FLAME initialization work
function[SimMat,CSOs,CluMat] = flame_init(XDat,KNN,N)
% FLAME_INIT calculates all initializationwork todo
%
% written by: Sandro Kolly
% date: 2011/08/18
M=zeros(N);
for l=1:N
for r=1:N
if l<r
M(l,r)=XDat(l,r)/(XDat(l,l)+XDat(r,r));
M(r,l)=XDat(r,l)/(XDat(l,l)+XDat(r,r));
end
end
end
for k=1:N
M(k,k)=0;
end
% sort matrix descending
SMat=zeros(N);
for im=1:N
S=sort(M(:,im),’descend’);
SMat(:,im)=S;
end
% take the sum of the KNN

SumList=sum(SMat(1:KNN,:));
AvSumList=SumList./KNN;
% take the longest of the KNN
DList=SMat(KNN,:);
% create adjacency matrix; if 1: adjacent

AdMat=zeros(N);
for ik=1:N
for in=1:N
if M(ik,in)>=DList(in) && ik ~= in;
AdMat(ik,in)=1;
end
end
end
% calculate the CSOs
CSO=ones(1,N);
AdSumMat=zeros(N);
for ad=1:N
AdSumMat(:,ad)=AvSumList’.*AdMat(:,ad);
end
CSOs=[];
for cs=1:N
for cso=1:N
if AvSumList(cs)<AdSumMat(cso,cs) && AdSumMat(cso,cs)~=0;
CSO(cs)=0;
Break;
end
end
if CSO(cs)==1
CSOs=[CSOs,cs];
end
end
%calculate the Outliers

%find CSO with smallest density
smallest=CSOs(1);
for el=CSOs
if AvSumList(el)<AvSumList(smallest)
smallest=el;
end
end
% Elements with less than 25% density of the smallest CSO are Outliers
Outliers=[];
for ol=1:N
if 10*AvSumList(ol)<AvSumList(smallest)
Outliers=[Outliers,ol];
end
end
% fill ClusterMatrix with Memberships:

% every Member equal to every CSO
% and every CSO full Membership to itself.
NumOfCSO=numel(CSOs);
NumOfOut=numel(Outliers);
CluMat=ones(N,NumOfCSO+1)/(NumOfCSO+1);
CSOVecs=eye(NumOfCSO+1);
for cmlc=1:NumOfCSO
CluMat(CSOs(cmlc),:)=CSOVecs(cmlc,:);
end
if NumOfOut>0
for cmlo=Outliers
CluMat(cmlo,:)=CSOVecs(NumOfCSO+1,:);
end
end
% create AdjacencyMatrix with distances InvMat=M.*AdMat;

% Similarity Matrix to get the weights from KNNs:
% InvMat=zeros(N);
% for sml=1:N
% for smr=1:N
% if AdDistMat(sml,smr)>0
% InvMat(sml,smr)=1/AdDistMat(sml,smr);
% end
% end
% end
SimList=sum(InvMat);
SimMat=zeros(N);
for sm=1:N
SimMat(sm,:)=InvMat(sm,:)./SimList;
end
return %% end of function flame_init()
References
[1] D. Arotaritei, S. Mitra, Web mining: a survey in the fuzzy framework, Fuzzy Sets Syst. 148 (2004) 5–19.
[2] R. Baeza-Yates, B. Ribeiro-Neto, Modern Information Retrieval: The Concepts and Technology Behind Search, 2nd ed., Addison-Wesley,
New York, 2011.
[3] A. Baker, Simplicity, [Online] Stanford Encyclopedia of Philosophy, Stanford University, California, 2010. Available at http://plato.stanford.
edu/entries/simplicity/ (accessed 10 August 2013).
[4] A. Beal, J. Strauss, Radically Transparent: Monitoring and Managing Reputations Online, 1st ed., John Wiley & Sons, 2008.
[5] C. Bizer, T. Heath, T. Berners-Lee, Linked data – the story so far, Int. J. Semantic Web Inf. Syst. 5 (3) (2009) 1–22.
[6] F. Bobillo, U. Straccia, Fuzzy ontology representation using OWL 2, Int. J. Approx. Reason. (20 May 2011), http://dx.doi.org/10.4018/jswis.
2009081901.
[7] P. Bourke, Intersection of two circles, 1999. [Online] Available at http://paulbourke.net/geometry/2circle/ (accessed 30 August 2012).
[8] D. Chandler, Semiotics: The Basics, 2nd ed., Routledge, 2007.
[9] A.K. Dey, G.D. Abowd, D. Salber, A conceptual framework and a toolkit for supporting the rapid prototyping of context-aware applications,
Hum.-Comput. Interact. 16 (2) (2001) 97–166.
[10] T. Ebert, Trust as the Key to Loyalty in Business-to-Consumer, Gabler Edition Wissenschaft, Wiesbaden, 2009.
[11] Franz, Franz Inc., 2011. [Online] Available at http://www.franz.com/agraph/allegrograph/ (accessed 30 August 2012).
[12] L. Fu, E. Medico, FLAME, a novel fuzzy clustering method for the analysis of DNA microarray data, BMC Bioinform. (4 January 2007),
http://dx.doi.org/10.1186/1471-2105-8-3.
[13] L. Gaines-Ross, Reputation warfare, Harv. Bus. Rev. (December 2010) 70–76.
[14] T.R. Gruber, A translation approach to portable ontology specifications, Knowl. Acquis. (1993), http://dx.doi.org/10.1006/knac.1993.1008.
[15] T.R. Gruber, Ontology of folksonomy: a mash-up of apples and oranges, 2005. [Online] Available at http://tomgruber.org/writing/ontology-of-
folksonomy.htm (accessed 30 August 2012).
[16] J.M. Hailpern, B.A. Huberman, Echo: the editor’s wisdom with the elegance of a magazine, in: Proceedings of the 5th ACM SIGCHI Sympo-
sium on Engineering Interactive Computing Systems, New York, NY, USA, 2013, pp. 313–322.
[17] Y. Hassan-Montero, V. Herrero-Solana, Improving tag-clouds as visual information retrieval interfaces, in: International Conference on Mul-
tidisciplinary Information Sciences and Technologies, Mérida, Spain, 2006.
[18] A. Hevner, S. Chatterjee, Design Research in Information Systems: Theory and Practice, Springer, Berlin, 2010.
[19] B. Huberman, F. Wu, The economics of attention: maximizing user value in information-rich environments, Adv. Complex Syst. (2008),
http://dx.doi.org/10.1145/1348599.1348602.
[20] E. Hüllermeier, Special issue on fuzzy sets in knowledge discovery, Fuzzy Sets Syst. 149 (2005) 1–3.
[21] D. Ingenhoff, U. Röttger, Issues Management. Ein Zentrales Verfahren der Unternehmenskommunikation, in: B.F. Schmid, B. Lyczek (Eds.),
Unternehmenskommunikation. Kommunikationsmanagement aus Sicht der Unternehmensführung, Wiesbaden, 2006, pp. 319–350.
[22] E. Kaufmann, Talking to the semantic Web, Dissertation, University of Zurich, Institute for Informatics, Zurich, 2009.
[23] M.A. Kaufmann, Inductive fuzzy classification in marketing analytics, Dissertation, University of Fribourg, Department of Informatics, Fri-
bourg, 2012.
[24] M.A. Kaufmann, E. Portmann, M. Fathi, A concept of semantics extraction from Web data by induction of fuzzy ontologies, in: IEEE
International Conference on Electro/Information Technology, Rapid City, SD, USA, 2013.
[25] J. Klewes, R. Wreschniok, Reputation Capital: Building and Maintaining Trust in the 21st Century, Springer, Berlin, 2010.
[26] B. Kosko, Fuzzy cognitive maps, Int. J. Man-Mach. Stud. 24 (1986) 65–75.
[27] G.A. Miller, WordNet: a lexical database for English, Commun. ACM 38 (11) (1995) 39–41.
[28] J. Nielsen, Discount usability for the Web. [Online] Jakob Nielsen’s Alertbox, 1997. Available at http://www.nngroup.com/articles/web-
discount-usability/ (accessed 10 August 2013).
[29] V. Novák, Reasoning about mathematical fuzzy logic and its future, Fuzzy Sets Syst. 192 (2012) 25–44.
[30] D.T. Parry, Fuzzy ontology and intelligent systems for discovery of useful medical information, Dissertation, Auckland University of Tech-
nology, 2005.
[31] W. Pedrycz, Fuzzy set technology in knowledge discovery, Fuzzy Sets Syst. 98 (3) (16 September 1998) 279–290.
[32] S. Pepper, Ontopedia, 2010. [Online] Available at http://www.ontopedia.net/pepper/papers/ELIS-TopicMaps.pdf (accessed 10 August 2013).
[33] E. Portmann, A fuzzy grassroots ontology for improving social semantic Web search, in: B. De Baets, R. Mesiar, L. Troiano (Eds.), Proceedings
of 6th International Summer School on Aggregation Operators, Benevento, 2011.
[34] E. Portmann, The FORA framework—a fuzzy grassroots ontology for online reputation management, Dissertation, University of Fribourg,
Department of Informatics, Fribourg, 2012.
[35] E. Portmann, M.A. Kaufmann, C. Graf, A distributed, semiotic-inductive, and human-oriented approach to web-scale knowledge retrieval, in:
Proceedings of the 21st ACM International Conference on Information and Knowledge Management, Maui, Hawaii, USA, 2012.
[36] E. Portmann, W. Pedrycz, Fuzzy Web knowledge aggregation, representation, and reasoning for online privacy and reputation management,
in: E. Papapgeorgiou (Ed.), Fuzzy Cognitive Maps for Applied Sciences and Engineering – From Fundamentals to Extensions and Learning
Algorithms, Intelligent Systems Reference Library, Springer, 2013.
[37] E. Portmann, A. Thiessen, Web 3.0 monitoring im stakeholder management, in: Andreas Meier, Marcel Blattner (Eds.), Web Monitoring, in:
HMD, vol. 293, dpunkt.verlag, Heidelberg, 2013, 50. Jahrgang.
[38] I. Robinson, J. Webber, E. Eifrem, Graph Databases, O’Reilly Media, Sebastopol, 2013.
[39] E. Sen, Social Media Monitoring für Unternehmen: Anforderungen an ein Web-Monitoring verstehen & die richtigen Fragen stellen, Social
Media Verlag, Norderstedt, 2011.
[40] H.A. Simon, Designing organizations for an information-rich world, in: Computers, Communications, and the Public Interest, Baltimore,
1971.
[41] J. Sterne, Social Media Metrics: How to Measure and Optimize Your Marketing Investment, John Wiley & Sons, Inc., Hoboken, New Jersey,
2011.
[42] T. Sudkamp, Examples, counterexamples, and measuring fuzzy associations, Fuzzy Sets Syst. 149 (2005) 57–71.
[43] E. Trillas, S. Guadarrama, What about fuzzy logic’s linguistic soundness? Fuzzy Sets Syst. 156 (2005) 334–340.
[44] P. Victor, C. Cornelis, M. De Cock, E. Herrera-Viedma, Practical aggregation operators for gradual trust and distrust, Fuzzy Sets Syst. 184
(2011) 126–147.
[45] K.E. Weick, K.M. Sutcliffe, Managing the Unexpected: Resilient Performance in an Age of Uncertainty, 2nd ed., Wiley, San Francisco, 2007.
[46] K. Weller, Knowledge Representation in the Social Semantic Web, de Gruyter Saur Verlag, Berlin/New York, 2010.
[47] F. Wu, B.A. Huberman, Novelty and collective attention, Proc. Natl. Acad. Sci. USA 104 (45) (Nov. 2007) 17599–17601.
[48] L.A. Zadeh, Fuzzy sets, Inf. Control 8 (3) (1965) 338–353.
[49] L.A. Zadeh, From computing with numbers to computing with words—from manipulation of measurements to manipulation of perceptions,
Int. J. Appl. Math. Comput. Sci. 12 (3) (2002) 307–324.
[50] L.A. Zadeh, BISC: the Berkeley initiative in soft computing, 2010. [Online] UC Berkeley, Available at http://www.cs.berkeley.edu/~zadeh/
presentations%202008/SSSC%2710-Full%20version-Precisiation%20of%20Meaning-July26.pdf (accessed 10 August 2013).

Portmann 2015

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Portmann 2015

Uploaded by

Copyright:

Available Formats

JID:FSS AID:6570 /FLA [m3SC+; v 1.194; Prn:7/07/2014; 16:20] P.

FORA – A fuzzy set based framework for online reputation

Fig. 1. Concept of knowledge map and its visualization.

2. Online reputation analysis

2.1. Measuring the reputation capital

Fig. 2. Online reputation management process.

2.2. The process of online reputation management

2.3. Applying fuzzy sets in online reputation management

Fig. 3. Reputation life cycle diagram.

3. The FORA framework

Fig. 4. The FORA framework architecture.

3.2. Reputation search engine

Fig. 5. Ontology-compilation process.

Algorithm 1 Metaphone Algorithm.

Algorithm 2 Plotting points.

Algorithm 3 FLAME Algorithm.

In the course of this process, we arrive at an approximation error,

Fig. 6. An exemplary fuzzy ontology schema.

Fig. 7. The reputation search engine.

3.3. Fuzzy grassroots ontology

SELECT ?apple WHERE

returning, by way of example (of Subsection 3.2), Granny Smith.

3.4. Dashboard with Topic Maps

Fig. 8. The FORA framework dashboard.

4. YouReputation prototype and lessons learned

Fig. 9. The YouReputation dashboard (from www.youreputation.org).

4.1. The YouReputation prototype

4.2. Apple Inc. use case

5. Conclusions and future directions

A.1. Comparison of distance measurements

Distance (continuous space) Distance (discrete space)

A.2. Comparison of fuzzy clustering algorithms

Factor Weight FCM GK FLAME

A.3. Comparison of knowledge administration systems

Factor Weight AllegroGraph Jena KAON

A.4. Comparison of knowledge representation systems

Factor Weight RDFS OWL Topic Maps Tag Clouds

A.5. Comparison of selected applications

Factor Sysomos MeMo Actionly YouReputation

B.1. Matlab calculation of the memberships of the clusters

Function [CluMat,turn,TagList,CSOs] = flame_calc(XDat,Tags,KNN,Turns)

[SimMat,CSOs,CluMat] = flame_init(XDat,KNN, N);

% calculate the Fuzzy Memberships by Local Approximation

B.2. Matlab FLAME initialization work

% take the sum of the KNN

% create adjacency matrix; if 1: adjacent

%calculate the Outliers

% fill ClusterMatrix with Memberships:

% create AdjacencyMatrix with distances InvMat=M.*AdMat;

return %% end of function flame_init()

You might also like