AN APPROACH TO PERSONALISATION IN E-LEARNING
SOCIAL ENVIRONMENTS
Hend Ben Hadji and Ho-Jin Choi
Korean Advanced Institute of Science and Technologies, Daejon, South Korea
Keywords: E-learning, Social tagging, Contextualization, Formal concept analysis, Similarity-based concept
approximation.
Abstract: We present an approach to a personalized learning service that takes advantage of social tagging to support
social learning in context and provides learning resources adapted to the abilities and needs of an individual
learner. We employ graph clustering technique in order to group tags into clusters having different contexts,
learner’s personomy to discover the learning context of the targeted learner, and Formal Concept hierarchy
theory to hierarchically cluster learning resources and retrieve the relevant resources to the learner’s needs.
1 INTRODUCTION
With the increasing growth of learning resources,
personalized support has become an essential
component of e-learning systems. Personalized
learning service aims at providing learning resources
adapted to the abilities and needs of an individual
learner. Several approaches in this direction have
been investigated. Some approaches employ
Semantic Web technologies to create semantic-rich
e-learning systems. They mostly rely on ontologies
to capture the semantics of the entire e-learning
process, including learning content, learner’s
characteristics, and learning context. Other
approaches propose social recommender systems for
recommending learning resources without the need
of understanding the semantics of the learning
content or the learners’ need. While other
approaches combine both technologies for better
personalization purpose.
Although all the diversity of approaches, e-
learning systems have not reached their full potential
in practice. This is because e-learning process tends
to undertake continuous changes over time, such as
changes to learning content or learner’s behavior.
Such changes are unfortunately not well supported
by prior approaches. (Vassileva, 2008) studied the
behaviour of the new generation of learners and
found that learners mostly learn by accessing to
social resource sharing systems, such as YouTube or
Del.cio.us, to find information, video, or any related
materials of interest. (Forte and Bruckman, 2008)
state that learning happens by consuming and
producing knowledge and provide the example of
the collaborative writing of Wiki articles as a
valuable learning experience. (Bateman, Brooks, and
McCall, 2006) propose a working prototype which
illustrate how socially constructed knowledge can
support domain experts in defining the semantics of
the learning process. Similarly, (Al-Kalifa and davis,
2007) demonstrate that metadata generated using
social tagging, core product of the social resource
sharing systems, is better than the metadata created
by human expert annotation in terms of search and
contextual coverage. Of all these works, no prior
work on learning personalization over Social Web
technologies has been explored yet. The main
approach being taking, here, focuses on leveraging
social tagging to support social learning in context.
We propose a personalized learning service which
can be built over any social resource sharing systems
and provides personalization capabilities.
The paper is structured as follows. Section 2
presents our motivation through a simple scenario of
learning over the Del.cio.us bookmarking system.
Section 3 details the phases of our learning
personalization process. Section 4 concludes the
paper and outlines the future work.
2 USE CASE SCENARIO
In this section we describe a simple e-learning
scenario over the social bookmarking system,
351
Ben Hadji H. and Choi H. (2010).
AN APPROACH TO PERSONALISATION IN E-LEARNING SOCIAL ENVIRONMENTS.
In Proceedings of the 2nd International Conference on Agents and Artificial Intelligence - Agents, pages 351-354
DOI: 10.5220/0002728103510354
Copyright
c
 SciTePress

Del.cio.us to motivate our approach. Consider a
student, Tom, is taking a Web Services course. Tom
has been given an assignment, which consists of
implementing a Web Service client using the SOAP
technology. As Tom wants learning by interacting
with other people and sharing knowledge about
topics of particular interest, he becomes an active
member of Del.cio.us; he uploads learning resources
that he is interested in and assigns them individual
tags for future retrieval. Further, Tom often
navigates in people’s space with the same interest
looking for relevant learning resources for making
assignments or reviewing exams.
As the users’ number accessing to Del.cio.us
increasingly grows, Tom finds it difficult this time to
find relevant resources about SOAP technology.
Therefore, he decides to issue a query looking for
learning resources which are tagged with “SOAP”.
As a result, thousand of resources that were
annotated with a SOAP tag are returned to Tom’s
query. Further, not all the documents are relevant to
the learner’s context. For example, a document
entitled “SOAP maker” was retrieved, and it is about
a shopping space for soap maker. This means that
SOAP tag has been used by users to refer to
different concepts in different contexts.
Figure 1: Social Tagging Mechanism.
Figure 1 illustrates such a situation where SOAP
tag is used by user 1 and user 2 to refer to
documents about Web Service, while used by user 3
to refer to Shopping context. To this end, we can say
that the Del.cio.us bookmarking system does not
implement an intelligent retrieval engine that tracks
the user’s behavior and returns relevant resources to
his/her context. This limitation constitute the main
barriers of adopting social resource sharing systems
as learning environments or even for implementing
social resource sharing system as a service of e-
learning systems. (Vassileva, 2008) believes that the
new generation of e-learning systems need to
support social learning in context, that is, support the
learner to find the right content and people while
offering social environments. In this context, we
propose a social learning personalization mechanism
that supports social learning and provides
personalization capabilities. Detailed description of
the process is in section 3.
3 PROPOSED ARCHITECTURE
Figure 2 presents the personalization process over
social learning environments. This process is
composed of three main phases: pre-processing
phase, contextualization phase, and retrieval phase.
3.1 Preprocessing Phase
This phase starts by collecting the learning resources
from the delicious website, number of users who
tagged the learning resources, and list of all tags
assigned to the learning resources. Once
folksonomies are collected, they are passed to the
preprocessing module which performs a series of
filters for cleaning the tags and removing ambiguous
and infrequent tags. Fist, tags are converted to lower
cases so that string manipulation can be performed.
Stop words and Non-English words are then
removed to ensure that only English tags are present
when recommending learning resources to the
learner. Finally, infrequent and isolated tags are
filtered out. At the end of the preprocessing phase,
we get concise tags, which are ready to be used in
the subsequent phases of the personalization process.
3.2 Contextualization Phase
This phase consists of two steps: (1) finding the
different contexts that might be related to the
learner’s query, and (2) identifying the suitable
context to the targeted learner. Let t be the tag query
and F’ the subset of the folksonomy that is
associated with that particular tag query defined as a
tuple F’ := (L
t
, R
t
, T
t
, A), where L
t
is the set of
learners who have used the tag t on one or more
learning resources: L
t
= {l |  r R,l, t ,r A }; R
t
the set of learning resources which have been
annotated with the tag t: R
t
= {r |  l L,l, t ,r 
A
}; and T
t
is the set of tags which have been used
together with t on some resources: T
t
= {t’ |  (l, r) 
L
×R ,l, t ,r A Λ l, t’ ,rA }. Having identified
the set of tags, T
t
, a weighted tag vector, v
r
, can be
constructed to represent a resource r, whose
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
352

elements correspond to the number of times a tag
has been assigned to it: v
r
= (v
r1
, v
r2
, v
r3
, …v
r|Tt|
). A
similarity matrix A = {a
ij
} is constructed to represent
the pairwise similarity of each resource by using
cosine similarity measure:
a
ij
= cos_similarity(v
r
,v
d
) (1)
A bipartite graph can be then constructed, where
|L
t
| vertices representing each of the resources and
edges weighted by the similarity between these
resources. This resource-based network might
display the tags used by the learner to refer to
different concepts in different contexts. For
example, a learner who is looking for resources
annotated with t = “SOAP” in Web Service context
might get resources referring to Shopping context.
Therefore, we should first discover the different
contexts in which t can be used and then find out the
context that matches the learner’s profile.
Figure 2: Pipeline Process of Social Learning
Personalization.
3.2.1 Graph Clustering
The process of finding out the different contexts in
which a tag query might be used is called Tag
Contextualization, term coined by (Yeung, Gibbins,
and al., 2009). This process consists of applying the
fast greedy algorithm to identify groups of vertices
in a network which are highly connected with those
in the same group of vertices but loosely connected
with those in other groups. The result of the tag
contextualization process is a set of k clusters:
C = {C
i
|1 ≤ i ≤ k}, where C
i
is a set of resources
clustered together to refer to a particular context.
Formally, each context is denoted as a triple
κ
i
=
(R
κi
, T
κi
, I), where R
κi
is the set of learning resources
in a cluster
C
i
: R
κi
= {r
i
|i=0, 1, … n & n ≤ |R
t
|}, T
κi
is the set of tags used within a cluster C
i
: T
κi
=
{t
j
|j=0, 1, … m & m ≤ |T
t
|} and I  R
κi
×T
κi
is a
binary relation defined between R
κi
and T
κi
, i.e.,
incidence matrix. Following FCA theory, the
elements of R
κi
are treated as objects, those of T
κi
as
the attributes of the objects collection. The
relationship between objects and attributes is
represented as cross tables, whose rows are headed
by the objects, whose columns are headed by the
attributes and whose cells are marked if the
incidence relation holds for the corresponding pair
of object and attribute.
3.2.2 Mapping Clusters to Learner’s Profile
Having identified the different contexts used with
the tag t, the next step is to find out the appropriate
context for the learner’s profile. The latter can be
obtained from the folksonomy which are associated
with that learner. This subset of the folksonomy is
called personomy. A personomy of a learner is
characterized as a tuple P
l
: = (R
l
, T
l
, A
l
), where R
l
is
the learner’s set of learning resources: R
l
= {r | t, r 
 A
l
}, T
l
is the learner’s set of tags: T
l
= {t| t, r  
A
l
}, and A
l
is the set of the tags of the user: A
l
= {(t,
r)|
l, t , r  A }. The learner profile can be then
described by two main set of elements, the resources
tagged and their associated tags. Finding the
appropriate context for the learner’s profile comes
then to find the similarity degree between his
personomy P
l
and each context κ
i
. A modified
Jaccard Index is used for similarity measure, defined
as follows:
(, ) ( )
ii
ii
lK lK
lj
lK lK
RR TT
Sim P K l
RR TT
αα
∩∩
=+−
∪∪
(2)
As shown in the above Equation, the Sim(P
l
,
κ
i
) function takes into account two factors, the tags
and resources. The similarity between tags reveals
the similarity in topics, while similarity between
documents reflects similarity in content. α is a
threshold, ranged between 0 and 1, which can be set
by the learner for flexibility. The greater similarity
value is the closet context to the learner’s profile is.
3.3 Retrieval Phase
Given the appropriate context of the learner, say κ =
(R
κ
, T
κ
, I), we need to find out the relevant
concept(s) in which the learner is interested. For
discovering the different concepts of a given context
and their relationships, we propose to use FCA
method because of its simplicity and effectiveness.
3.3.1 Tags Conceptualization
Central to FCA is the notion of formal concepts. A
formal concept of formal context
κ is defined as a
pair (R
j
, T
j
), where R
j
 R
κ
, T
j
 T
κ
, R
j
’ = T
j
, and R
j
’
= T
j
’. R
j
’ is a set of attributes common to the objects
in R
j
which is defined as R
j
’ = {t  T
κ
 r  R
j
: (r,
t)  I} and T
j
’ is the set of objects commonly have
the attributes in T
j
which is defined as T
j
’ = {r  R
κ

AN APPROACH TO PERSONALISATION IN E-LEARNING SOCIAL ENVIRONMENTS
353

t  T
j
: (r, t)  I}. Accordingly, (R
j
, T
j
) is a formal
concept if the set of tags shared by the learning
resources in R
j
is identical with T
j
and on the other
hand R
j
is also the set of all learning resources in the
collection having all attributes in T
j
. R
j
is then called
the extent and T
j
the intent of the formal concept (R
j
,
T
j
). For the sake of simplification, the formal
concept (R
j
, T
j
) is denoted C
j
, T
j
is Intent(C
j
), and R
j
is the Extent(C
j
).
3.3.2 Concept Approximation
The final step consists of finding out the closet
conceptual cluster to the targeted learner and
retrieving the relevant learning resources belonging
to it. Learning resources are retrieved here using
Concept approximation method, proposed by
(Saquer and Deogun 2001). This method consists of
computing the similarity degree between a set of
attributes (i.e., tags added by the learner in our case)
to be approximated and the formal concepts on the
given context, and select the most similar formal
concept approximation result. The similarity
measure is defined as (Saquer, Deogun, 2001):
Intent( ) ( ) Extent( )
Intent( ) ( ) Extent( )
()
2
C
BCB C
BCB C
fB
α
α
∩∩
+
∪∪
=
(3)
The range of f
C
(*) is the interval [0,1]. For the
attribute set T
i
, f
C
(T
j
) = 0 when T
j
and (T
j
) are
disjoint from the intent and extent of C, respectively.
F
C
(T
j
)= 1 when T
j
= Intent(C) and, therefore, (T
j
)
= Extent (C). In general the closer the value f
C
(T
j
) to
1, the greater the similarity between T
j
and the intent
of C. To approximate a set of tags S, we find a
formal concept C that maximizes the value of f
C
(S).
In case more than one formal concept are found to
approximate S with same value of f
C
, we say that
these concepts equally approximate S.
4 CONCLUSIONS
In the light of the advances of the Social Web
technologies, we believe that there is a potential for
enhancement of e-learning systems. In this paper, we
present our approach to social learning in context.
This approach can be applied to any social resource
sharing systems which store learning resources and
enable the production of broad folksonomies, but
also to any e-learning systems which integrate social
tagging service. The future work consists of
implementing the whole process over the
bookmarking system Del.cio.us and comparing
results with the classical retrieval approach.
ACKNOWLEDGEMENTS
This research was supported by the MKE(Ministry
of Knowledge Economy), Korea, under the
ITRC(Information Technology Research Center)
support program supervised by the NIPA(National
IT Industry Promotion Agency) (NIPA-2009-
(C1090-0902-0032)).
REFERENCES
Al-Khalifa, H.S., Davis, H.C. (2007). Replacing the
Monolithic LOM: A folksonomic Approach. In
ICALT’07, 7
th
IEEE International Conference on
Advanced Learning Technologies.
Bateman, S., Brooks, C.A., and McCall, G. Collaborative
Tagging Approaches for Ontological Metadata in
Adaptive Learning Systems. In Proc. SWEL at AH’06.
(2006) 3-12.
Yeung, C.A., Gibbins, N., Shadbolt, N. (2009).
Contextualising tags in collaborative tagging systems.
In the Proceedings of the 20
th
ACM Conference on
Hypertext and hypermedia, Italy.
Forte, A., Bruckman, A. (2008). Scaling consensus:
increasing decentralization in Wikipedia governance.
In the Proceedings of Hawaiian International
Conference of Systems Sciences (HICSS).
Mika, P. (2005). Ontologies Are Us: A Unified Model of
Social Networks and Semantics. In Proc. ISWC 2005.
Saquer, J., Deogun, J. (2001). Concept approximations
based on rough sets and similarity measures. Journal
of Applied Mathematics and Computer Science, 11
(2001), pp. 655–674.
Vassileva J., (2008), “Towards Social Learning
Environments”. IEEE Transactions on Learning
Technologies, Vol. 1 No. 4, October-December, pp.
199-214, IEEE CS & ES.
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
354