Preference Dissemination by Sharing Viewpoints

Simulating Serendipity

Guillaume Surroca

, Philippe Lemoisson

, Clément Jonquet

and Stefano Cerri

Laboratoty of Informatics, Robotics Micro-Electronics of Montpellier, Montpellier, France

UMR Territories, Environment, Remote Sensing and Spatial Information, CIRAD, Montpellier, France

Keywords: Knowledge Representation, Knowledge Discovery and Dissemination, Serendipity, User-Centered

Knowledge Engineering, Collective Intelligence, Web 2.0, Agents.

Abstract: The Web currently stores two types of content. These contents include linked data from the semantic Web

and user contributions from the social Web. Our aim is to represent simplified aspects of these contents

within a unified topological model and to harvest the benefits of integrating both content types in order to

prompt collective learning and knowledge discovery. In particular, we wish to capture the phenomenon of

Serendipity (i.e., incidental learning) using a subjective knowledge representation formalism, in which

several “viewpoints” are individually interpretable from a knowledge graph. We prove our own Viewpoints

approach by evidencing the collective learning capacity enabled by our approach. To that effect, we build a

simulation that disseminates knowledge with linked data and user contributions, similar to the way the Web

is formed. Using a behavioral model configured to represent various Web navigation strategies, we seek to

optimize the distribution of preference systems. Our results outline the most appropriate strategies for

incidental learning, bringing us closer to understanding and modeling the processes involved in Serendipity.

An implementation of the Viewpoints formalism kernel is available. The underlying Viewpoints model

allows us to abstract and generalize our current proof of concept for the indexing of any type of data set.

1 INTRODUCTION

Since Web 2.0 has democratized the sharing,

recommendation and creation of content via social

networks, blogs and fora, and since semantic Web

technologies have begun to structure the knowledge

deposited, generated and stored on the Web, two

kinds of content have emerged. These types of

content differ in the ways they are produced and

structured. On one hand, contribution-based social

Web platforms allow the production of a wealth of

data with little or no structure; these data evolve

rapidly (e.g., folksonomies (Mika, 2007)). On the

other hand, highly structured knowledge is

constituted consensually by circles of experts (e.g.,

ontologies (Karapiperis and Apostolou, 2006) or

linked data (Bizer et al., 2009)). With the

Viewpoints approach, our objective is to create a

knowledge representation formalism that retains the

best qualities of each type of content. Our objective

is to support and give value to both (i) the structure

which characterizes semantic Web datasets and

(ii) the evolution and maintenance rates of shared

knowledge on the social Web as proposed in Gruber

(Gruber, 2008) or (Freddo and Tacla, 2009). We aim

to contribute to knowledge representation

approaches by designing a system involving Web

agents (human or artificial) who share “viewpoints”

linking system resources (identified by a URI). We

ask ourselves the following questions:

 Which Web browsing strategies allow the most

optimal diffusion of user preference systems?

 What should the conditions be to favor incidental

learning, a.k.a., Serendipity, in the study of

preference systems?

We define the preference system of an agent by the

expression of his tastes and attractions in terms of

proximity or distance relationships between Web

resources. In a previous contribution (Lemoisson et

al., 2013), we demonstrated the learning ability of a

knowledge base built with an initial version of our

formalism. However, this proof of concept was

based on a poor behavioral model of agents who

navigated and contributed to the knowledge base; we

were only interested in the agents' satisfaction and

did not take into account their preference systems. In

402

Surroca, G., Lemoisson, P., Jonquet, C. and Cerri, S..

Preference Dissemination by Sharing Viewpoints - Simulating Serendipity.

In Proceedings of the 7th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2015) - Volume 2: KEOD, pages 402-409

ISBN: 978-989-758-158-8

another contribution, we showed how Viewpoints

allow the search and discovery of knowledge

through a search engine prototype for scientific

publications (Surroca et al., 2014). In the newest

model, we include a “Serendipity acceptance” factor

in the behavior of agents, defined as the tendency of

an agent to turn to resources outside of his

preference system. This allows us to assess the

diffusion of preference systems, depending on

whether an agent is open-minded or focused on what

he knows and prefers. Using this model, we build a

simulation based on individual behavior rules

(microscopic level) in order to observe the effect on

collective learning and on the diffusion of preference

systems (macroscopic level). This simulation

illustrates the advantages of using Viewpoints to

“merge” the essence of data semantics and the social

Web.

The rest of this article is organized as follows:

section 2 presents the background and inspiration for

our approach by introducing the notion of

Serendipity in computer systems. In our review of

the state of the art, we also briefly compare

Viewpoints to several other knowledge

representation approaches. Then, we briefly present

the Viewpoints formalism in Section 3. Section 4

explains our behavioral model of Web users and our

representation of their preference systems: we show

how we simulate the evolution of the Web as a

knowledge graph and discuss a set of hypotheses on

the impact of individual browsing strategies. Section

5 presents a simulation in which three agents (the

Princes of Serendip) contribute to building a 'toy'

knowledge graph with resources of different shapes,

sizes and colors; then we discuss our current results

relative to our assumptions and our research

objectives. Section 6 concludes and presents

potential perspectives for this work.

2 STATE OF THE ART

2.1 Knowledge Representation

Several studies have focused on the merging of the

Semantic Web and the Social Web (Gruber, 2008).

We synthetically compare our approach to these

studies as follows: in addition to incorporating the

(human or artificial) Agent as presented in (Mika,

2007), our representation of knowledge considers it

a central constituent. We explain how in the

formalism section. Moreover, our knowledge

representation considers Viewpoints micro-

expressions of individual semantics. However, our

mechanism for evaluating and confronting

Viewpoints does not use any additional contribution

as is the case in (Limpens and Gandon, 2011). Thus,

the emphasis is placed on what emerges from the

knowledge graph, as reported in (Aberer et al., 2004;

Noh et al., 2010); indeed, these authors studied the

possibility of the emergence of a collective

representation of knowledge with a "bottom-up"

vision of system interactions.

Finally, we define a metric distance over the set

of resources formed by the knowledge providers

(Agents), supports (Documents) and descriptors

(Topics) while semantic distances found in the

literature apply to homogeneous subclasses such as

distances between tags or ontology concepts (Lee et

al., 2008). The resulting Viewpoints Knowledge

Graph (KG) is constituted by resources connected by

viewpoints, and can be seen as a wide, evolving,

associative memory enabling collective intelligence,

metaphorically replicating a brain, where all learning

processes are supported by the evolving strength of

synapses (Edelman, 1987). Instead, we adopt a

topological approach and compute semantic

distances on top of the Viewpoints in a manner

similar to (Pedersen et al., 2007).

2.2 Serendipity, the Incidental

Learning

The term ‘Serendipity’ is derived from an ancient

Persian tale entitled ‘The Three Princes of Serendip’

(Merton and Barber, 2006). Recently, Perriault said

that "the Serendipity effect (...) consists in nimbly

and randomly happening upon something we did not

search for". We are then led to make abductive

inferences in order to build a theoretical framework

which encompasses, via appropriate aggregation,

information which used to be disparate (Perriault,

2000). We note that the notion of luck or chance is

important in the Serendipity phenomenon. However,

"it does not only depend on a divine dice roll" as

explained in (Fine and Deegan, 1996) and takes

place only at the border of what is already known.

Thus, incidental learning is greatly facilitated when

new knowledge is in the vicinity of existing

knowledge and may be interpreted by someone who

knows this neighborhood. We share the vision that

knowledge does not guarantee serendipitous

discovery, but that it makes it more likely. We

therefore introduce the notion of Serendipity

proximal zone, which is similar to the concept of

proximal development zone (Vygotsky, 1978) in

learning and education sciences. We will show

below how the Serendipity acceptance factor helps

Preference Dissemination by Sharing Viewpoints - Simulating Serendipity

403

us to capture Serendipity in our model.

When considering the huge amount of

information available on the Web and the ways in

which one may get lost while browsing, Serendipity

seems to be a realistic phenomenon. One may talk

about serendipitous Web-based learning, as

explained hereafter. The search for knowledge

through serendipitous learning can succeed by

chance or as an offside activity of a main task

(Bowles, 2004). For instance, a user who makes an

initial query may be progressively led into an

unexpected path that ultimately proves more

productive than the initial search. In such cases,

Bowles writes that serendipitous learning occurs

(Bowles, 2004). This is exactly the phenomenon we

model and observe in our section 4 with multiple

navigation strategies. In addition, according to Allen

Tough, almost 80% of learning is informal and

unplanned (Tough, 1999). Serendipitous navigation

is an "intellectual lottery (...) with small chances but

with big potential payoff" (Marchionini, 1997). In

the latter work, the parallel with our Viewpoints

approach is made explicit: "We also gain new

viewpoints and associations for our problem by

browsing alternative sources using different tools,

techniques and data structures."

Recommender systems (Adomavicius and

Tuzhilin, 2005) are increasingly interested in

Serendipity, because the variety of recommendations

is as important as their accuracy. Serendipity goes

beyond what recommendation systems offer, thanks

to the surprise, variety and novelty of the proposed

results. Additionally, many recommender systems

have begun to implement Serendipity principles. The

folksonomy-based recommendation in (Yamaba et

al., 2013) allows users to tag books and go beyond

the traditional classification, and therefore add new

books to the Serendipity proximal zone of other

users. However, to our knowledge, except from

work proposed in (Corneli et al., 2014) on the

theoretical framework for the phenomenon of

Serendipity, the literature on the formalization and

the measurement of this phenomenon is lacking.

Based on our review, there is currently no

exploitable model of Serendipity.

3 VIEWPOINTS FORMALISM

Viewpoints is a formalism for subjective knowledge;

it holds that any proximity or distance relationship

between two resources is expressed by an agent as a

viewpoint. A typed viewpoint connects these two

resources. These viewpoints are individually

interpreted by a perspective chosen by the user /

contributor. This perspective allows assigning a

weight to each viewpoint, depending on who issued

it, on when it was created, and on its semantic type

or other more complex criteria. Therefore,

Viewpoints is a knowledge representation formalism

centered on equally considered human (e.g., Web

users) or artificial (e.g., data mining tools,

knowledge extractors, ontologies) agents. Resources

(providers, descriptors and knowledge supports) are

bound by the viewpoints on the knowledge graph.

The KG is a bipartite graph consisting of a set of

resources R and a set of viewpoints V connecting

these resources. The resources in R are either agents

(knowledge providers, i.e., viewpoint creators),

knowledge descriptors (topics, tags) or knowledge

supports (documents, videos, Web pages, messages,

posts, etc.). A viewpoint is a tuple

(a → {r1, r2}, θ, t) containing the following

information:

 a, the agent who issued the viewpoint;

 {r1, r2}, the couple of resources semantically

connected by a;

 θ, the viewpoint's type, used to interpret (i.e.,

assign a weight to) it;

 t, the viewpoint's creation date.

For instance, (Guillaume → {Diffusion systems [...]

views, acm:Knowledge representation and

reasoning}, dc:subject, 27/02/15) means that the

agent Guillaume associates this article to the

Knowledge representation and the reasoning concept

of ACM’s taxonomy with the relation DublinCore

subject. (Mario → {Mario, Luigi}, foaf:knows,

1985) means that Mario elicited that he has known

(as in FOAF) Luigi since 1985. To identify the

meaning of the data represented in the form of

Viewpoints, we adopt, when possible, existing

Semantic Web types.

4 VIEWPOINTS EXPLOITATION

The aggregation of all connections between two

resources created by the different agents form a

semantic proximity link named synapse. The

strength of the synapse is based on the aggregation

of the weights of each viewpoint in the synapse. The

two functions of evaluation (Map) and aggregation

(Reduce) of viewpoints form a perspective which

allows the exploitation of subjective knowledge. For

the same KG, several interpretations, defined as

Knowledge Maps (KM), can be made dependent on

how agents evaluate and aggregate viewpoints. The

KEOD 2015 - 7th International Conference on Knowledge Engineering and Ontology Development

404

Knowledge Map is a graph made of resources (R)

and synapses (S) to which common graph algorithms

can be easily applied. The perspective is unique to

each user, who decides to interpret the KG any way

he wants. The two functions of evaluation and

aggregation of viewpoints can be extended at will to

suitably match one’s needs. Figure 1 illustrates the

interpretation process of KG. In the following

simulation we use: (i) a direct neighborhood

function that returns all the resources directly

connected by viewpoints to a specified resource, and

the weight of the synapses binding this resource to

its direct neighbors; (ii) an indirect neighborhood

function based on the Dijkstra algorithm and which,

for a resource ri, returns all resources rj on all the

paths starting from ri with a length less than a

specified threshold, m (simulation-specific

parameter).

Figure 1: Interpretation of Knowledge Graph (KG) into

Knowledge map (KM).

An important aspect, directly inspired from the

Web 2.0, lies in the built-in feature for integrating

agent feedback. Within their perspective, agents use

any Viewpoint for browsing KM and reversely

update the KG through viewpoints expressing their

feedback. Along these exploitation/feedback cycles,

shared knowledge is continuously elicited against

the beliefs of the agents in a selection process. The

knowledge map is defined as a graph in which

semantic similarities within the knowledge resources

are computed according to a given perspective.

5 SERENDIP SIMULATION

Our goal is to simulate the evolution of a knowledge

base – such as the Web – from individual behavior

rules that describe agents browsing the Web and

disseminating their preference systems. First, we

explain how we represent the preference systems in

a Viewpoints KG; then, we propose a behavioral

model simulating different configurable navigation

strategies. This model is based on calculations of

direct and indirect neighborhoods. Finally, we

observe the effect of this set of individual rules on

the macroscopic level of knowledge represented in

the resulting KG.

5.1 Preference Systems Representation

In this simulation, each resource is characterized by

a shape, a size and a color. Shape and size

information will already be included in the KG at the

beginning of the simulation; this information

simulates the Semantic Web data. Color information

is introduced step-by-step during the simulation by

three agents, the princes of Serendip, who know and

like a different color each (red, green, blue); this

information simulates social Web contributions. The

preference system of a prince is the set of all the

viewpoints he has issued to make same-color

resources get closer to him or closer to one another.

We consider two kinds of viewpoints: (i) the first

kind links two same-color resources (vps:knows ;

(ii) the second associates a prince of a specific color

with a resource of the same color (vps:likes). The

dissemination of a preference system is therefore

equivalent to the distribution of the color

information in the graph, i.e., the more colored the

graph becomes, the more a preference system has

been shared. Thus, when the graph “learns” a color,

it illustrates the collective intelligence of the

community.

For example, if the red prince searches a red

resource r and retrieves a red-color resource r', he

issues the two following viewpoints (RedPrince →

{Redprince, r},vps:likes,τ) and (RedPrince → {r,r'},

vps:knows, τ). In the next section we will present the

different navigation strategies which allow princes

to disseminate knowledge about their color.

5.2 Behavioral Model of the Serendip

Princes

The state automaton in Figure 2 describes the

behavior of the princes when they are navigating in

the KG and disseminating their preference systems

(viewpoints emission). More generally, this

automaton simulates the behavior of a user when he

is exploring the contents of a knowledge base such

as the Web.

Preference Dissemination by Sharing Viewpoints - Simulating Serendipity

405

Figure 2: Behavioral automaton of the Princes of

Serendip.

We capture behaviors such as: querying a search

engine, exploring the results, following links

included in these results and querying the search

engine again. In our simulation the behavior of a

Prince corresponds to a specific configuration of the

β, μ and σ parameters; we call this a navigation

strategy. Our simulation is divided into cycles that

correspond to successive explorations of the KG. At

the beginning of a cycle, a prince begins interacting

with the KG; we simulate the use of a search engine:

A resource of the KG is randomly selected and the

indirect neighborhood function is used to retrieve a

list of results (other resources) sorted by semantic

proximity. From the proposed results, the prince

continues (low β) or abandons this search and

undertakes a new one (high β). If he continues, he

must evaluate these results one by one (comparing

them to the color corresponding to his preference

system) and select the first non-visited result based

on the σ parameter. If the prince accepts Serendipity

(high σ), he does not systematically select resources

of his own color; if he does not accept Serendipity

(low σ), he will instead focus on resources of his

color only. Once a resource is chosen, the prince

moves to the next stage of his journey: Depending

on μ, he will either perform a direct search on this

resource (high μ) or explore locally around this

resource (low μ). The first interaction simulates the

act of opening a Web page as the result of a previous

search; the second interaction simulates either a new

search, e.g., with the title or content of the current

page, or clicking on a Web link within a page. In the

simulation, princes start with an initial budget of

interactions; this budget is decreased with each

interaction (research and exploration). It represents

the amount of effort princes are willing to make

when navigating. When princes wish to go

backwards, three scenarios will lead to the end of the

cycle: There are no previous steps; or, all resources

have been visited; or, the initial interaction budget

has been spent.

These strategies simulate Web browsing. In

terms of graph traversal, a high β corresponds to a

breadth-first approach, whereas lower β corresponds

to a depth-first approach. In an information search

process, the breadth-first approach would

superficially assess all the results and get an overall

idea of all the results; instead, the depth-first

approach would rather focus on what would seem to

be the best result and dig deeper. μ determines the

navigation style. A high μ value means princes

mainly use SEARCH engines that sort results

according to a global approach; a low μ means

princes will carry out a step-by-step exploration by

collecting unsorted local results (EXPLORATION).

For example, navigating from one suggested

YouTube video to another is a good illustration of a

step-by-step exploration, in which as a succession of

Google searches illustrate a BREADTH traversal.

We represent the Serendipity acceptance factor (σ)

as a third dimension. High σ means princes are

mainly OPEN and are willing to visit both the

resources that match their preferences and the

resources that do not but could lead to chance

discoveries. Low σ means princes are mainly

CLOSED to the latter prospect and are entirely

guided by their preferences when browsing.

6 SIMULATION DYNAMICS

6.1 Initial Conditions

A fixed-size KG is generated. In addition to their

specific color (red, green, blue), the resources of the

KG are characterized by their size (small, medium,

large) and their shape (square, circle, triangle). For

each possible size, shape and color combination, N

resources are created. Therefore, there are initially

27N resources. Two artificial agents, called peons

are added to the KG. One of them shares his

appreciation of shapes in the knowledge graph,

connecting all the same shapes of resource pairs by

viewpoint types vps:initial. The other peon does the

same for size. Thus, after the peons have shared their

appreciations, the KG does not “know” colors

because resources are only tied by size and shape

characteristics. Finally, the 3 princes are added to

the KG. Each of them is characterized by a unique

color, and has the ability to appreciate colors and

share this assessment by issuing new viewpoints

such as vps:like and vps:knows in the KG. Thus,

there is an implicit understanding that the princes are

KEOD 2015 - 7th International Conference on Knowledge Engineering and Ontology Development

406

only able to share by issuing viewpoints as feedback.

6.2 Dissemination of Preference

Systems

The simulation parameters are summarized in Table

1. The princes follow the behavioral model

previously described and disseminate their

preferences (knowledge of their own color) by

issuing

vps:like

and

vps:knows

viewpoints. The

weight assigned to each type of viewpoint is shown

in Table 1. The aggregation capability of viewpoints

for the calculation of the value is the sum of the

synapses. At the end of each cycle, the following

measures are calculated to evaluate the

dissemination of color (preference) knowledge in the

KG:

 M1 Color X: This is the ratio of the average

distance between any resources over the average

distance between X-colored resources.

 M2 Color X: This is the probability of getting a

resource of the same color in the neighborhood

of X-colored resource.

Table 1: Simulation parameters.

Categories Parameters

Values

(if fixed)

Scale parameters

Scale factor (N) 3

Number of cycles 100

Number of iterations per

cycle

Perspective

parameters

Weight of viewpoints with

type vps:initial

… type vps:knows

… type vps:like

Navigation

strategy

parameters

ẞ

Activity

distribution

Red prince 33% 80%

Green prince 33% 10%

Given the large number of parameters (Table 1),

we present the results (curves) of several simulations

with the parameter configurations which we

consider the most significant for navigation

strategies. However, we explain the effects of

specific parameters in the discussion section. Other

fixed parameter values are given in Table 1.

6.3 Hypotheses

Princes progressively share their color assessments

with other users through the feedback mechanism.

We aim to observe how the KG "learns" (at the

global level) the notion of color that was not

originally in the knowledge represented by the

vps:initial viewpoints. Thanks to viewpoints, each

individual preference system becomes part of the

collective knowledge represented in the KG, where

it coexists with the preference systems of other

princes. Our goal is to experiment with different

navigation strategies and demonstrate that

preference systems do not neutralize each other

when concurrently broadcast. We also want to

measure the effect of Serendipity. Thus, we expect

M1 to increase; in other words, the average distance

between same-color resources will decrease more

quickly than the average distance between any

resource. M2 should also increase as it reflects the

probability of finding the same-color resource in the

m-neighborhood of a resource.

7 RESULTS AND DISCUSSIONS

7.1 Impact of the Serendipity

Acceptance Factor

We start by assessing the impact of σ on the

dissemination of the color red thanks to measures

M1 and M2 Red. One can notice (Figure 3) that

when search engines are mainly used, M1 and M2

increase at a faster rate when Serendipity acceptance

is low; conversely, when Serendipity acceptance is

high, they reach higher final values. Therefore,

Serendipity acceptance allows a wider dissemination

of color knowledge. Indeed, while the search

indirectly returns results and allows the creation of

viewpoints that have not already been issued,

Serendipity acceptance increases the potential for

creating new original viewpoints.

Figure 3: M1 Red and M2 Red evolution when princes

mainly use search engines (µ=70%, ẞ=10%).

Preference Dissemination by Sharing Viewpoints - Simulating Serendipity

407

These new associations are expressions of

preference systems that would likely not have been

generated if the princes had been guided only by

their preferences to navigate. In contrast, we observe

(Figure 4) that when mainly local exploration results

are used to navigate from, Serendipity acceptance

does not affect either M1 and M2 value increases or

final values. This strategy’s idea is to explore local

and in-depth results; moreover, going through less

interesting results along the road tends to slow the

spread of preference systems. The μ effect

(navigation device) is very important for

Serendipity. However, we realize that the relative

homogeneity of our graph does not realistically

represent the Web’s structure. We believe that,

under more realistic conditions, Serendipity can

produce more substantial gains than it does in our

"toy" knowledge graph. In this simulation, the three

princes are active (33%) and ẞ = 10%.

Figure 4: M1 Red and M2 Red evolution when princes

mainly use link-by-link exploration (µ=30%, ẞ=10%).

7.2 Adaptation to Real Web Data

We also conducted a similar experiment with real

data on movies and user ratings. We studied a Web

dataset (MovieLens), in which explicit semantics

were mixed with social contributions. This dataset

consisted in two sets of 100,000 and 1,000,000

ratings which had been collected by the GroupLens

Research Project at the University of Minnesota. In

our MovieLens experiment, users elicited

preferences when they associated movies with

ratings. Initially, each movie was linked to other

movies by metadata such as actors, directors or

genres. For instance, the genre characteristic

corresponded to the shape characteristic in our

Serendip simulation. All films, as well as other

resources such as genres, were initially added to the

KG. During each cycle, a portion of the ratings was

added to the KG as viewpoints, once again

simulating the contributions of the social Web. We

observed knowledge crystallizing progressively

around the reviewers. This experiment showed us

that when working with such a recommendation

system, we may observe that structured data (genres,

actors, director) do bootstrap the creation of

subjective (social) knowledge. Integrating user data

such as gender, age group, job and movie metadata

(genre, release year) showed us new relations. User

was closing movies and movies were semantically

reproaching users. One of the goal we gave to us

with ViewpointS was also to observe dynamics in an

evolving represented knowledge.

8 CONCLUSIONS AND

PERSPECTIVES

After presenting and positioning our approach of

subjective knowledge representation, we studied the

phenomenon of Serendipity and its current influence

on the Web. With the Princes of Serendip

simulation, we presented an experiment for

modeling Serendipity on the Web. We recognize

that this behavioral model of Web users may not

fully represent the reality and diversity of Web

exploration methods. Nonetheless, we hope that we

have demonstrated the ability of the Viewpoints

knowledge graph to learn. Our simulation results

allowed us to assess the contribution of the

Serendipity acceptance factor to various navigation

strategies and its impact on the dissemination of

preference systems; we consolidated the Viewpoints

proof of concept by confronting it with a more

realistic use of modeling and simulation. We are

planning for several applications which may help us

evaluate the Viewpoints approach: Amongst them,

(i) one will consist in cross scientific discovery of

agronomic knowledge (CIRAD) and (ii) another will

deal with biomedical data within the SIFR project

(http://www.lirmm.fr/sifr). We are finishing also

several IR benchmarks (recall, precision and f-

mesure) on a film recommendation scenario

comparing our semantic neighborhood methods to

classic indexation and research methods such as

Vector Space Model. We will soon publish

benchmarks results in one the scenarios we

previously mentioned.

KEOD 2015 - 7th International Conference on Knowledge Engineering and Ontology Development

408

ACKNOWLEDGEMENTS

This work was supported in part by the French

National Research Agency under JCJC program,

grant ANR-12-JS02-01001, as well as by University

of Montpellier, CNRS and the CIRAD.

REFERENCES

Aberer, K., Cudr, P., Catarci, T., Hacid, M., Illarramendi,

A., Mecella, M., … Scannapieco, M. (2004).

Emergent Semantics Principles and Issues. In D. Lee,

YoonJoon and Li, Jianzhong and Whang, Kyu-Young

and Lee (Ed.), Database Systems for Advanced

Applications (Vol. 2, pp. 25–38). Springer Berlin

Heidelber.

Adomavicius, G., & Tuzhilin, A. (2005). Toward the next

generation of recommender systems: a survey of the

state-of-the-art and possible extensions. IEEE

Transactions on Knowledge and Data Engineering,

17(6), 734–749.

Bizer, C., Health, T., & Berners-Lee, T. (2009). Linked

Data - The Story So Far. In Semantic Services,

Interoperability and Web Applications: Emerging

Concepts (Vol. 5, pp. 1–22).

Bowles, M. (2004). Relearning to E-learn: Strategies for

Electronic Learning and Knowledge. Educational

Technology & Society, 7(4), 212–220.

Corneli, J., Pease, A., & Colton, S. (2014). Modelling

serendipity in a computational context. arXiv Preprint

arXiv:1411.0440.

Edelman, G. (1987). Neural Darwinism: The theory of

neuronal group selection.

Fine, G. A., & Deegan, J. G. (1996). Three principles of

Serendip: insight, chance, and discovery in qualitative

research. International Journal of Qualitative Studies

in Education, 9(4), 434–447.

Freddo, A. R., & Tacla, C. A. (2009). Integrating social

web with semantic web : ontology learning and

ontology evolution from folksonomies. KEOD 2009

Proceedings, 247–253.

Gruber, T. (2008). Collective knowledge systems: Where

the Social Web meets the Semantic Web. Web

Semantics: Science, Services and Agents on the World

Wide Web, 6(1), 4–13.

Karapiperis, S., & Apostolou, D. (2006). Consensus

building in collaborative ontology engineering

processes. Journal of Universal Knowledge

Management, 199–216.

Lee, W.-N., Shah, N., Sundlass, K., & Musen, M. (2008).

Comparison of ontology-based semantic-similarity

measures. AMIA, Annual Symposium 2008, 384–8.

Lemoisson, P., Surroca, G., & Cerri, S. (2013).

Viewpoints : an alternative approach toward Business

Intelligence. In eChallenges e-2013 (p. 8).

Limpens, F., & Gandon, F. (2011). Un cycle de vie

complet pour l ’ enrichissement sémantique des

folksonomies. In Extraction Gestion de Connaissance

EGC 2011 (pp. 389–400).

Marchionini, G. (1997). Information Seeking in Electronic

Environments (Cambridge., p. 224). Cambridge

university press.

Merton, R. K., & Barber, E. (2006). The Travels and

Adventures of Serendipity: A Study in Sociological

Semantics and the Sociology of Science (Princeton.,

Vol. 2006, p. 313).

Mika, P. (2007). Ontologies are us: A unified model of

social networks and semantics. Web Semantics:

Science, Services and Agents on the World Wide

Web, 5(1), 5–15. doi:10.1016/j.websem.2006.11.002

Noh, T., Park, S., Park, S., & Lee, S. (2010). Learning the

emergent knowledge from annotated blog postings.

Web Semantics: Science, Services and Agents on the

World Wide Web, 8(4), 329–339.

Pedersen, T., Pakhomov, S. V. S., Patwardhan, S., &

Chute, C. G. (2007). Measures of semantic similarity

and relatedness in the biomedical domain. Journal of

Biomedical Informatics, 40(3), 288–99.

Perriault, J. (2000). Effet diligence, effet serendip et autres

défis pour les sciences de l’information. In Pratiques

collectives distribuées sur Internet.

Surroca, G., Lemoisson, P., Jonquet, C., & Cerri, S. A.

(2014, May 13). Construction et évolution de

connaissances par confrontation de points de vue :

prototype pour la recherche d’information scientifique.

IC - 25èmes Journées Francophones d’Ingénierie Des

Connaissances.

Tough, A. (1999). Reflections on the Study of Adult

Learning. WALL Working Paper.

Vygotsky, L. S. (1978). Mind in Society: The

Development of Higher Psychological Processes (Vol.

1978, p. 159).

Yamaba, H., Tanoue, M., & Takatsuka, K. (2013). On a

serendipity-oriented recommender system based on

folksonomy. Procedia Computer Science, 22, 276–

284.

Preference Dissemination by Sharing Viewpoints - Simulating Serendipity

409