A CONTACT RECOMMENDER SYSTEM FOR A MEDIATED
SOCIAL MEDIA
Michel Plu, Layda Agosto
France Telecom R&D; 2 avenue Pierre Marzin Lannion France
Laurence Vignollet; Jean-Charles Marty
Laboratoire SYSCOM, Unniversité de Savoie, Campus scientifique,Le Bourget Du lac Chambery France
Keywords: Collaborative filtering, recommender systems, social network analysis, online communities, social media
Abstract: Within corporate intranet or on the WWW, a global search engine is the main service used to discover and
sort information. Nevertheless, even the most "intelligent" ones have great difficulties to select those
targeted to each user specific needs and preferences. We have built a mediated social media named
SoMeONe, which helps people to control their information exchanges through trusted relationships. A key
component of this system is a contact recommender, which helps people to open their relationship networks
by exchanging targeted information with qualified new users. Instead of using only matching between
interests of users, this "socially aware" recommender system also takes into account existing relationships in
the social network of the system. In this paper, we describe the computations of those recommendations
based on a social network analysis.
1 A NEW MEDIA FOR
PERSONALIZED ACCESS TO
INFORMATION
A large part of companies' knowledge is embedded
in each employee's documents. Web technologies
are now being used to make those numerous
documents easily accessible through a decentralized
intranet or extranet. The WWW also provides access
to many interesting resources to any employees but
they are lost through the huge quantity of available
pages. Those information networks are becoming
essential for being correctly informed. However, in
such a web environment, information is distributed
throughout the company or through the WWW. This
makes it difficult to find information which is useful
and relevant to each user’s needs.
One of the great challenges of search engine
tools, mainly based on an artificial (computer-based)
centralized intelligence, is to be able to select
relevant answers according to user's preferences,
background, or current activity., … In order to face
this personalization challenge, we are developing a
complementary approach based on users' distributed
intelligence where the relevancy of a resource for a
user is based on the existing references to this
resource from other users and the trustworthiness of
relationships between those users. This is based on
our assumption that some users might prefer to trust
other users than machine to obtain good advice
about information resources. We are thus
introducing a user-centric approach as opposed to a
computer-centric one to develop a new intelligent
interface for accessing the WWW.
This approach is supported by our collaborative
system named SoMeONe (Social Media using
Opinions through a trust Network) (Agosto, 2003).
This system is particularly adapted to users
preferring to access information which already has a
certain approval, for instance, information coming
from appreciated or skilled people in corresponding
domains.
Key issues in this system are motivating users to
exchange information and helping them to manage
and optimise their relationship network. To deal
with those problems we have integrated in SoMeOne
a contact recommender system, which suggests that
107
Plu M., Agosto L., Vignollet L. and Marty J. (2004).
A CONTACT RECOMMENDER SYSTEM FOR A MEDIATED SOCIAL MEDIA.
In Proceedings of the Sixth International Conference on Enterprise Information Systems, pages 107-114
DOI: 10.5220/0002636801070114
Copyright
c
 SciTePress

some users exchange information with new users.
We make the assumption that users will be
motivated to produce and exchange good
information in order to be recommended by the
recommender. Those recommendations are not only
based on the common interests of users but also on
social qualities of each user. This "socially aware
recommender system" is the focus of this paper.
2 SOCIAL MEDIA
The idea of using communication networks as a
support tool to find "focused" people is not new.
Newsgroups and mailing lists are the most famous
examples of such collaborative systems. By using
them, people are acquiring a new, social, cyber-
behaviour that asks them to adopt new habits in
working and even in thinking schemas. They form
online communities in the sense of J. Preece
(Preece, 2000). We call "social media" systems
capable of relating persons to establish relationships.
We call "mediated social network" the social
network of a social media.
Using information technology can help to
improve the flow of pertinent information between
people and the global efficiency of the system by
analysing the structure of a mediated social network.
Such a mediated social network can be used to
receive very personalized recommendations of
information resources carefully selected by trusted
users. By doing this, we develop a new vision where
information navigates from users to users instead of
having users navigating through information. We
named this vision the "web of people" (Plu, 2003).
The ultimate goal is to help people to get in contact
with appropriate persons according to the diversity
of their needs to find and filter suitable information.
Let's now look more deeply into one of the key
issues presented before: the user motivation to share
information. We assume this requirement to be true.
Indeed, we believe that in our information society,
and more particularly in a competitive and dynamic
business environment, this collaborative behaviour
is crucial for an awareness of new information and
in order to receive support or credits from others.
Bourdieu and others have also largely demonstrated
the value of social capital not only as being the
knowledge of individual workers but also the
relations between them (Bourdieu, 1986).
Consequently, it is sensible for companies that want
to develop their social capital to develop and support
cooperative behaviour in the everyday practice of
their employees.
But even if this collaborative behaviour is
supposed to be natural for our users, it has to be
applied to our system. To deal with this requirement,
one can imagine having a regulation component,
which organizes the behaviour of users and applies a
user management policy (Durand, 2003). An
alternative approach is to integrate some
components in the system to influence such users'
behaviour in order to have them following the
required behaviour rules. To illustrate how a
technology can influence user's behaviour, one can
look to how indexing technologies used by major
Internet search engines have transformed the way
web authors are designing their web pages.
The contact recommender system we are
presenting is such a component. Within the
SoMeONe system, a user has to be recommended to
be able to receive information from new users. Thus,
the recommender can recommend users with the
required social behavior. However, having
interesting information might not be sufficient for
being recommended. The recommender has also to
analyse the defined social qualities of the users'
participation into the mediated social network. These
social qualities of a user can depend for example on
the credits s/he receives from others or the
originality of his/her contribution (which means that
no user could replace his/her contribution). One can
imagine many other social qualities to qualify the
user willingness to collaborate and the value or
his/her participation to the community. Those social
qualities can be computed using social network
analysis techniques (Wasserman, 1994).
We call "socially aware recommender system"
a recommender system that takes into account those
social qualities to compute and rank its
recommendations.
3 SOMEONE: A COOPERATIVE
SYSTEM FOR PERSONALIZED
INFORMATION EXCHANGE
To experiment those ideas, we have integrated such
a recommender in our SoMeONE system (Agosto,
2003). The main goal of this system is to support
the creation and management of mediated social
networks. It helps users to exchange
recommendations about good contents available
through an information network like the WWW or
corporate intranet. It is supposed to help people to
improve and to optimise their mediated social
network in order to discover and find information
resources, which are adapted to their needs, taste,
background, culture or any other personal features
which make humans so different.
ICEIS 2004 - HUMAN-COMPUTER INTERACTION
108

The way to share personal information in
SoMeONe is described as follows:
– Each user manages a personal taxonomy, in
order to annotate and to index their documents.
Each element in that taxonomy is called a topic.
A document could be for instance an email, an
image, a video, or a report. In fact, it is anything
that can be identified with an URL.
– When displayed, all information associated with
a document (also called meta-information) is
aggregated. For that, we introduce the concept of
review. Reviews are created by associating
topic(s) and other information (like a text
annotation) on documents.
– The accessibility of reviewed information, and
thus the exchange of information between users,
depends on the accessibility of topics in the
reviews. The accessibility of a topic is defined
according to a list managed by the topic owner;
this list is called a topic distribution list (TDL for
short). It groups the users allowed to access all
information having a review with the topic.
– We call a user's contacts, the set of users
belonging to the distribution list of at least one of
his/her topics. Those contacts could be friends,
colleagues, family members, or any others.
Information is exchanged between users when
they access the system using their personal home
page. This page lets the user navigates through all
information s/he is allowed to access, and let
him/her to create new reviews for personal indexing
purposes. However, creating a new review to a
document discovered from a received review on that
document makes it accessible to all the new users in
the TDL of the topics associated to the new review.
In consequence, personal indexing is automatically
associated to information forwarding. As a result,
information in the reviews, including document
references, flow through the network of users
according to the topic's TDL. We called "semantic
addressing", this information routing process based
on the indexing of information. This is the basic
principle of the "web of people" where information
navigates from users to users instead of having users
navigating through information (Plu, 2003).
4 A "SOCIALLY AWARE"
RECOMMENDER SYSTEM
The recommender we have developed and integrated
in SoMeONe lets people have new contacts. It
suggests to a user to add some users to the
distribution list of some topics.
For this, the recommender needs first to identify
topics which show the similar interests of two users.
Like many others do, our recommender system is
also using a collaborative filtering approach
(Resnick, 1997). The originality of our work lies in
the fact that we complement this approach with the
computation of new ranking features based on social
network analysis (Wasserman, 1994). The goal is to
filter the recommendations obtained from the
collaborative filtering process according to a
personal information requirement and users social
qualities corresponding to it. We qualify such a
recommender as "socially aware".
In a social network analysis, people, groups or
organizations that are members of social systems are
treated as "sets of nodes" (linked by edges) –forming
networks. They represent social structures. Given a
set of nodes, there are several strategies for deciding
how to collect measurements on the relations among
them. Matrices or vectors can be used to represent
information, and algebraic computations are done to
identify specific patterns of ties among social nodes
(Wasserman, 1994).
Differences in how users are connected can be a
key indicator of the efficiency and "complexity" of
the global social organization supported by the
mediated social network. Individual users may have
many or few ties. Individuals may be "sources" of
ties, "sinks" (actors that receive ties, but don't send
them), or both. The analysis of the relations between
users can indicates a degree of "reciprocity" and
"transitivity" which can be interpreted, for instance,
as important indicators of stability.
The graph structure analysis of a mediated social
network can be used for many purposes. It might be
used to show users' roles, their position, their global
appreciation, their dependency to communities to
which they belong. It is also useful in order to
qualify the exchanged information. Further in this
paper, we present how we use these analysis
techniques to propose new contacts.
Furthermore, social network analysis has also
been largely used in a sub-field of classical
information retrieval called biblio-metrics to analyse
citations in scientific papers (Garfield, 1972). It has
also led to the development of new algorithms for
information retrieval algorithms for hypertext like
PageRank (Brin, 1998). They are mainly based on
the computation of a centrality measure of the nodes
in a graph formed by web pages. The assumption is
that a link provides some credit to the linked page
The social network we extract from the
mediated social network supported by SoMeONe, is
a directed graph consisting of a set of nodes with
directed edges between pairs of nodes. Nodes are the
topics of users and edges are their relations. Those
relations between two topics are computed
A CONTACT RECOMMENDER SYSTEM FOR A MEDIATED SOCIAL MEDIA
109

according to reviews being associated within those
two topics. Thus, in this social network, there is an
edge i from a topic v to a topic u, if the owner of
topic u is receiving and taking information
associated to topic v. In other words, the owner of
topic u is in the distribution list of the topic v and
takes at least one review containing the topic v and
creates a new review on the same document with
his/her topic u. Consequently, the graph
representation will show the relation v → u.
The relation v → u indicates the flow of
appreciated information through the network. It
means that the owner of topic u is receiving and
appreciates information from the owner of topic v.
Figure 1: Mediated social network example
Figure 1 shows a graphical representation of a
small part of such a network. In this example, there
is six users. Each box shown as folders represents
some of the topics of these users. Each relation v →
u between topics is presented by a directed lattice.
Reviewed information resources are noted with a
lower case letter and a number. A label on a lattice
means that a resource has been discovered from a
review in the source topic.
Our socially aware recommender system first
takes into account the interest of users and then takes
into account the state of the users topics in the social
networks.
In the first step, it finds the relationships of users
with approximate interests (not only commons
ones). This means that for instance, we avoid giving
only recommendations directly obtained from
intersections of appreciated items in the users'
profiles, which is generally the strategy of existing
systems. This first feature is obtained by our
collaborative filtering techniques using repositories
of already classified items (Plu, 2003). Second, the
user can control the type of contact
recommendations s/he is going to receive. This
means that a user can define the strategy to rank
computed recommendations. This last feature is
accomplished by our SocialRank algorithm, which
completes our collaborative filtering algorithm.
The SocialRank algorithm uses some social
properties to filter topics which are candidates for
recommendations (those topics initially computed
with the collaborative filtering algorithm). The
social properties used depend on the information
strategy chosen by the users. They are computed by
using the SoMeONe's social network described
above.
By using those social properties as filters, two
users with the same interest would not receive the
same recommendations of contacts. Thus, this
should avoid the traditional problem of "preferential
attachment" in network based communication
systems (Adar, 2000). The preferential attachment
problem rises when most of users communicate with
the same very small group of users. Recommending
only experts to everyone could lead to this situation.
We will see below (see section 5.3.4) how
SoMeONe prevents such a situation by letting users
choose another information strategy than the
"Looking for Experts" strategy. More generally,
different "social properties" computed from the
social network analysis can be used to choose the
contact recommendations in order to influence the
way the social network will evolve! Thus, a socially
aware recommender system can help to give the
social network some interesting global properties
depending on the global criteria the designer of a
social media wants to optimise. Such interesting
properties can be, for instance: a good clustering
factor, a small diameter, a good global reciprocity
or/and transitivity factor.
 Online
communities
b4, b5
 Web
technologies
+
a2,
+
b5,
+
g1,
+
g2
 Java
a1, a2, f1
 New
technologies
g1, g2, g3,
+
b5,
+
f1
 Objects
+
a2
 Developing
+
a2
f1
b5
g1, g2
a2
a2
a2
b5
 Internet
g4
 Online
communities
b4, b5
 Web
technologies
+
a2,
+
b5,
+
g1,
+
g2
 Java
a1, a2, f1
 New
technologies
g1, g2, g3,
+
b5,
+
f1
 Objects
+
a2
 Developing
+
a2
f1
b5
g1, g2
a2
a2
a2
b5
 Internet
g4
Jean-Charles
Laurence
Layda
John
Michel
Layda
Pascal
 Online
communities
b4, b5
 Web
technologies
+
a2,
+
b5,
+
g1,
+
g2
 Java
a1, a2, f1
 New
technologies
g1, g2, g3,
+
b5,
+
f1
 Objects
+
a2
 Developing
+
a2
f1
b5
g1, g2
a2
a2
a2
b5
 Internet
g4
 Online
communities
b4, b5
 Web
technologies
+
a2,
+
b5,
+
g1,
+
g2
 Java
a1, a2, f1
 New
technologies
g1, g2, g3,
+
b5,
+
f1
 Objects
+
a2
 Developing
+
a2
f1
b5
g1, g2
a2
a2
a2
b5
 Internet
g4
Jean-Charles
Laurence
Layda
John
Michel
Layda
Pascal
We assume that some users will be seeking to be
recommended to others. Therefore, by using some
specific social properties in the recommendation
process, we think the recommender system can
influence the motivation and participation of the
users. In other words, if users know the strategy used
by the recommender system, we can assume that
some users will try to adapt their behaviour
according to it.
To be able to test this idea, we have first
implemented the computation of some social
properties and we have implemented some
information strategies using those properties in order
to select appropriate contact recommendations.
In order to let users to select one of the
implemented strategies which best fit their needs we
have ascribed “names” and descriptions to them.
Here are the three we have already implemented and
experimented:
– "Looking for Experts". The user only trust
credited experts who filter information for him.
– "Gathering all". The user want to have the
widest coverage of a topic, thus gathering as
much information as possible,
ICEIS 2004 - HUMAN-COMPUTER INTERACTION
110

– "Going to the sources". The user wants to
obtain the newest information rapidly, avoiding
users who are acting as intermediaries.
We have started with these three strategies but
our goal is to look for new ones or improving the
existing ones. By default, the "Going to the source"
strategy is selected, but users can change it by
editing her/his personal profile. This choice can be
refined for each personal topic.
The formulae related to the computation of the
social properties used by each "strategy" are
explained in the SocialRank section.
5 COMPUTING CONTACT
RECOMMENDATIONS
In this section we are going to present the three steps
of our recommendation process. Firstly, we describe
the collaborative filtering algorithm used to compute
potential contact recommendations based on topic
similarities using an existing classification of a large
amount of URLs. Secondly, we specify the
computation of social properties of each topic in the
SoMeONe’s social network. Finally, we show how
we filter the potential contact recommendations
obtained in the first step according to the topic
similarities, the social properties of the
recommended topics, and the information strategy
chosen by users.
5.1 Collaborative filtering
The bases of the collaborative filtering algorithm
that we have built are presented in (Plu, 2003). It
uses URL co-citations analysis. Co-citation is
established when two users associate personal
reviews to the same documents or to different
documents referenced within the same category of a
WWW directory. The recommendations of contacts
are computed using one or more specialized
directories. By directories, we mean repositories of
web sites categorized by subject. For our tests we
have started with the one provided by the Open
Directory Project (http://www.dmoz.org).
The collaborative filtering algorithm (CFA)
computes similarity between topics. It has to detect
the case of two topics having reviews with URLs
equal or similar to the URLs classified in the same
ODP category. The CFA computes a similarity
measure between each topic and each ODP category.
Like others do, this similarity measure is based on
URLs co-citation analysis. (the URLs to which the
reviews inside topics make reference). This
similarity measure is computed according to the
formula given in (Plu, 2003).
The CFA only computes the similarity between
topics that do not belong to the same user. Pairs of
similar topics noted (t1, t2) for topics labelled t1 and
t2, are sorted according to the similarity measure S.
Contact recommendations are then computed from
those similar topics.
5.2 SocialRank
The SocialRank algorithm filters the topic
recommendations according to some of their social
properties.
Having the topics' taxonomy of users, and the
distribution list of the topics defined, we are able to
extract the social network explained above. We
model this directed graph as an adjacent matrix.
Each matrix element represents the relationship
between two topics. As introduced above, a
relationship is established when a user creates new
reviews from other reviews received from other
users. They thus establish relationships between
their topics within the created reviews and the topics
of others within the received reviews. To take into
account the importance of each relation, each vertex
is weighted with a measure W(e,f) representing the
number of documents received from topic f and then
reviewed with a topic e. We compute a matrix W
with each element noted W(e, f), topic e being in the
row and topic f in the column of the matrix, for the
vertex from f . W(e,f) is computed with the formula:
or W(e, f) = 0 if card(e)=0
(1)
)(
),(*
),(
ecard
feCard
feW =
Card*(e,f) counts all the documents having a
review with the topic e and a review with the topic f,
the review with topic f being older than the review
with topic e; card (e) is the total number of reviews
with topic e.
Using this W matrix, the SocialRank algorithm
also computes one square matrix and two vectors of
topics:
– A vector of experts E, in order to obtain the
expert topics.
– A redundancy matrix R, in order to obtain
redundant topics.
– A vector of originals O, in order to obtain
original topics.
The computation of these matrix and vectors could
be obtained by different methods, as clearly
explained in (Wasserman, 1994).
To identify topics as "experts" we use a common
centrality measure of a topic defined recursively
according to the centrality of the topics receiving
A CONTACT RECOMMENDER SYSTEM FOR A MEDIATED SOCIAL MEDIA
111

information from it. Each element E(e) of the expert
vector is defined according to the recursive formula:
(2)
∑
= hEehWeE )(*),()(
For the computation of vector E we use the
algorithm named PageRank and used for WWW
pages (Brin, 1998). But the matrix used has to reflect
a reputation relation ("e is giving reputation to f",
f←e). We consider that this relation is the invert of
the relation modelled in our matrix W, which
reflects the flow of information through the topics
(f→e). Indeed, if a user reviews documents received
with topic f with his topic e, then topic e is giving
reputation (credit) to topic f. That is why we use the
weight W(h, e) instead of W(e, h) to compute E(e).
The PageRank algorithm requires that the
weights of the adjacent matrix W(e, f) have to be
modified in W*(e, f) in order to have the following
needed convergence properties (see (Brin, 1998) for
more details). This is partly achieved because the
new weights W*(e, f), once normalized, represent
the probability for a document being reviewed with
topic f to be reviewed with a topic e. Thus, our
matrix W corresponds to a stochastic matrix.
Following the PageRank algorithm, we also
complete the graph with new connections in order to
have all nodes connected.
To compute redundancy and originality, we
first define vectors G(e) as the set of all topics g
connected to topic e. Second, we define P(e, f) as the
proportion of the relation between topic e and f
among all the relations with topic e. P(e, f) is
computed with the formula:
If else P(e,f)=0 (3)
)(eGf ∈
The evaluation of redundancy between topics is
computed in a matrix R. We define that a topic e is
redundant with f if both are the same type of
information sources because they have the same
information obtained from the same sources.
Explicitly, the redundancy between e and f depends
on:
– If f is connected with e. This means that e is
receiving information from f.
– If topics connected to e are also connected to f.
This means that topics sending information to e
are also sending it to f.
We compute R(e, f) according to the following
formula:
(4)
Finally we compute the vector O to represent
original topics. The originality of a topic is measured
according to the novelty of URLs in the topic
compared to the URLs received from connected
topics. A topic e is original if it contains more URLs
discovered by the owner of the topic than received
from other topics. It also depends on the number of
URLs in the topic. We compute the vector O
according to the following formula:
∈Hh
(5)
∑
∈
−=
)'
),(1)(
eGh
heWeO
5.3 Applying SocialRank
Now we illustrate these calculations with our social
network example presented in figure 1 where there
are six actors, seven topics shown as folders, and
reviews noted with a lower case letter and a number.
The URLs of the reviews belong to 4 ODP
categories noted A,B,F,G. For example we note "a1"
a review having an URL referenced in the category
A of the ODP directory. A label on a lattice means
that a URL has been discovered from a review in the
source topic.
In this example, we suppose that the user Layda
wants to obtain recommendations about her topic
Internet. The CFA similarities computation produces
the following recommendations: (Internet → New
technologies) and (Internet → Web technologies)
because those three topics have reviews on URLs
referenced in the category G of the ODP category
(even if their intersection is empty). A
recommendation noted (t1→t2) means that owner of
the topic t2 should be in the distribution list of the
topic t1 if it is not the case.
Those initial recommendations are going to be
analysed by our SocialRank algorithm. One issue of
the analysis is which topic the system will
recommend to Layda related to her topic Internet,
Web technologies or New technologies (or both)? R
is an important matrix because it helps to decide if
two topics are redundant to each other. If so, which
of them are more relevant to recommend according
to the user specific needs? This decision is going to
be applied to the topics Web technologies (noted
WT) and New technologies (Noted NT).
Before the computation of R, we first have to
compute W and P. From (1) we compute W(WT,
NT). Then, we have:
(we assume that b5 were reviewed by WT before being
reviewed by NT).
This means that the average of information
received by Web technologies from New
technologies is 0.75, which is high (meaning that
their relation is important).
∑
)
, f
fP
∈
=
)(
,(
)(
),(
eGg
gew
eW
e
75.0
4
3
)(
),(*
NT) W(WT, ===
WTcard
NTWTCard
()
(, ) (, ) (, ) ( , )
gGe
Re f pe f peg p f g
∈
=+
∑
ICEIS 2004 - HUMAN-COMPUTER INTERACTION
112

Here are the matrix W and P for our example:
P NT WT Java OC
NT 0.5 0.5
WT 0.6 0.2 0.2
With matrix P, we obtain the proportion of the
relation between WT and NT among all the relations
with WT. The value 0,6 indicates an important
relation between both topics.
5.3.1 Evaluating redundant topics
As we explained above, matrix R helps to decide if
two topics are redundant to each other. From (4),
R(WT, NT) can be computed as
This value indicates a redundancy between WT
and NT, which reveals that WT could be a similar
information source to NT; therefore, it is relevant to
recommend only one of them.
The same computation gives R(NT,WT) =
0,2.Notice that R(WT,NT) > R(NT,WT) ! This is an
important result because it helps the system to
decide which topics to recommend according to the
user's strategy. We will develop this in a later
section.
5.3.2 Evaluating experts
Let's now compute the expert property. If we follow
(2), we will obtain E(WT) =0.095879; E(NT)=
0.080576 for topics WT and NT. This result is
interpreted as follows:
– Web technologies is the more expert topic. We
can notice (figure 1) that even if it does not have
its own reviews, it has collected different
reviews from two topics having a good level of
expertise. Web technologies is supplying with its
information two other topics, Objects and
Developing, who are giving to it a kind of
credibility or reputation.
– New technologies is at second level of expertise
From figure 1, we can see that it has collected
different reviews from two topics with a good
level of expertise but it is supplying only one
topic with its information! Remember that the
computation of E is based on a centrality
measure indicating a reputation degree (Brin,
1998). However, its level of expertise being
higher than a defined threshold this topic is kept
as candidate for being recommended.
W NT WT Java OC
NT 0.2 0.2
WT 0.75 0.25 0.25
5.3.3 Evaluating original topics
By applying (5), we obtain the next O vector values:
Topic O(e)
Internet 1.0
Java 1.0
Online Communities 1.0
New technologies 0.6
Web technologies -0.25
Developing 0.0
Objects 0.0
The result is interpreted as follows:
– Internet is the more original topic. The
originality of Internet is evident because it is
isolated, because it is not redundant with the
others and because it can bring new information.
Java and Online communities are also original
topics because URLs have been reviewed with
them before the other topics (see figure 1).
8.0
),(p),(p
),(p),(p
),(p),(p
),(pNT) R(WT, =
⎥
⎥
⎥
⎦
⎤
⎢
⎢
⎢
⎣
⎡
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎝
⎛
+
+
+=
NTNTNTWT
JavaNTJavaWT
OCNTOCWT
NTWT
– However, comparing their place in the vector O,
NT is more original than WT.
5.3.4 Applying users' strategies
Because WT and NT have been identified as
redundant, only one will be chosen according to
Layda's information strategy. If she has selected:
1. Looking for experts: This leads to the
selection of a topic with the highest Expert property;
the answer of the recommender would be WT.
2. Gathering all: The answer with this strategy
is the topic having the highest value for R, therefore
it would be WT because R(WT,NT) > R(NT,WT)
(reinforcing the global approval of WT over NT).
3. Going to the sources: the selected topic
would be NT, because the strategy gives priority to
the most originals among topics with a sufficient
level of expertise.
What happens if Layda does not define an initial
strategy? We explained that one of the priorities of
our mediated system is avoiding the preferential
attachment problem (Jin, 2001). Therefore, the
default strategy is "Going to the sources", because it
should improve the reactivity of the social networks
by minimizing intermediaries. Another important
situation to encourage is the connection of
independent components.
In order to protect user's information privacy, no
user can add his identifier to the topic access list of
any other user's private topics. Thus,
recommendations displayed only suggest sending
information to new users. In our example, the
A CONTACT RECOMMENDER SYSTEM FOR A MEDIATED SOCIAL MEDIA
113

system will recommend to Layda to add Michel
owner of NT or Laurence, owner of WT to the
distribution list of her topic Internet. But we assume
that a user receiving new information will also send
back new information. To encourage such reciprocal
relationships the recommender needs also to check if
the topic Internet satisfies Michel's or Laurence's
information strategy for their topic NT or WT. Thus
finally the recommender will try to choose the topic
that will stratify the best the strategy of the two users
involved in the suggested relationship.
6 CONCLUSION
In this paper, we've proposed to improve an original
information exchange system, SoMeONe, which
facilitates the creation of relationships between
users, in order to cover each user's information need.
We've included a contact recommendation module
that helps users to open their closed relational
network and thus discover new sources of
information.
We had proposed in (Plu, 2003) to use a
collaborative filtering algorithm. This algorithm
suggests that a user exchanges reviews on
information source that they have already evaluated
or produced. But these recommendations have to be
carefully chosen in order to not let him/her having a
too big relational network and for the global
efficiency of the social media. Thus, our SocialRank
algorithm presented in this article filters those
recommendations using the computation of one
matrix and two vectors. This lets the system propose
to users several information strategies to establish
new relationships.
Many recommender systems have already been
studied and some of them are operational like online
bookshops (Resnick, 1997). However, our system
recommends users instead of recommending
contents. Thus it is more similar to McDonald's
expertise recommender (McDonald, 1998). But as
far as we know, none of the recommender systems
integrate a traditional collaborative filtering
algorithm with social properties resulting from
social network analysis. The use of social network
analysis to improve information retrieval in
enterprise is also recommended in (Raghavan,
2002). But this paper does not present any
recommender system in order to establish exchange
relationships between users. Our work was partly
inspired by the ReferalWeb system (Kautz, 1997)
but in our system, we've introduced social properties
and the social network is manually controlled by
users, and evolves according to users accepting
contact recommendations.
In order to test our ideas, we've introduced the
system in the Intranet of France Telecom R&D and
in the portal of the University of Savoie, inside the
project called "Cartable Electronique"®. The usage
of our system in these different contexts should
allow us to validate our initial hypothesis: a
recommendation process of carefully selected
contacts should incite users to produce interesting
information and develop collaborative behaviour.
REFERENCES
Agosto L., Plu M. Vignollet L, and Bellec P., 2003
SOMEONE: A cooperative system for personalized
information exchange In ICEIS’03,Volume 4, p. 71-78,
Kluwer.
Adar E. And Huberman B., 2000, Free Riding on
Gnutella. First Monday, 5(10). (www.firstmonday.dk).
Bourdieu, P.,1986. The forms of capital. In J. Richardson
(Ed.), Handbook of theory and research for the
sociology of education (pp. 241-258). New York:
Greenwood.
Brin S. and Page L, 1998, The anatomy of a large-scale
hypertextual (Web) search engine. In The Seventh
International World Wide Web Conference.
Durand G., Vignollet L., 2003, Améliorer l’engagement
dans un collecticiel", to be published IHM 2003, Caen,
France, November 2003.
Emily M. Jin, Michelle Girvan, and M. E. J. Newman,
2001, The structure of growing social networks, Phys.
Rev. E 64, 046132.
Garfield E, 1972, Citation analysis as a tool In journal
evaluation. Science, 178.
Kautz, H., Selman B., Shah M, 1997, Referral Web:
Combining Social Networks and Collaborative
Filtering. CACM 40(3): 63-65.
McDonald, D. and Ackerman, M.S., 1998. Just talk to me:
A field study of expertise location. In Proc. CSCW’98,
pp. 315-324. New York: ACM.
Plu M., Agosto L., Bellec P., Van De Velde W. 2003, The
Web of People: A dual view on the WWW, In Proc.of
The Twelfth International World Wide Web
Conference, Budapest, Best Alternate Track Paper.
Raghavan, P., 2002, Social Networks: from the Web to the
Enterprise. IEEE Internet Computing, pp. 91-94,
Jan/Feb 2002.
Preece, J. 2000, Online Communities: Designing
Usability, Supporting Sociability. Chichester, UK:
John Wiley & Sons, 439 pages.
Resnick, P, and Varian, 1997, H.R. Recommender
Systems. Commun. ACM 40, 3 (56-58).
Wasserman, S. and Faust K., 1994, Social Network
Analysis: Methods and Applications. Cambridge
University Press.
ICEIS 2004 - HUMAN-COMPUTER INTERACTION
114