Semantic Collaborative Filtering for Learning Objects
Recommendation
Lamia Berkani
1, 2
and Omar Nouali
3
1
Departement of Computer Science, USTHB University, Bab-Ezzouar, Algiers, Algeria
2
Higher National School of Computer Science (ESI, ex. INI), Oued Smar, Algiers, Algeria
3
Department of Research Computing, CERIST, Algiers, Algeria
Keywords: Online Communities of Practice, Learning Resources, Recommendation System, Ontologies, Semantic
Filtering, Collaborative Filtering.
Abstract: The present paper proposes a personalized recommendation approach of learning objects (LOs) within an
online Community of Practice (CoP). Three strategies of recommendation have been proposed: (1) a
semantic filtering (SemF) by member’s interests; (2) a collaborative filtering (CF) based on the member’s
expertise level; and (3) a semantic collaborative filtering combining in different ways the two approaches.
The expertise level of a member is calculated in relation to all of his domains of expertise using the domain
knowledge ontology (DKOnto). A similarity measure is proposed based on a set of rules which cover all the
possible cases for the relative positions of two domains in DKOnto. In order to illustrate our work, some
preliminary results of experimentation have been presented.
1 INTRODUCTION
The great expansion and explosive use of the
Internet has created new ways of collaboration
between people as well as exchange and sharing of
knowledge. A vast number of learning object
repositories are made available to any user searching
for educational content on various topics
(Tzikopoulos et al., 2007). However, one of the
main problems encountered actually is the selection
of the appropriate resources. Accordingly, to deal
with the problem of information overload, the need
for recommender systems is more than necessary.
The main objective of our research is to facilitate
access and reuse of knowledge within a CoP of
teachers. The main objective of this community is to
promote e-learning in higher education context
applied to the domain of computer science.
We propose in this paper a personnalized
recommendation approach of learning objects (LOs)
for members of this CoP, based on the semantic
collaborative information filtering approach. Three
strategies of recommendation have been proposed:
(1) a semantic filtering (SemF) by member’s
interests; (2) a collaborative filtering (CF) based on
the member’s expertise; and (3) a semantic
collaborative filtering combining in different ways
the two approaches. These strategies are based
respectively on the following member’s objective:
specialization ; learning; or both, specialization and
learning. The CF is used to predict the utility of LOs
for members based on the similarity among their
preferences and the preferences of other members.
The SemF is used, to take advantage of the enhanced
semantics representation.
The main contribution of this paper concerns: (1)
the proposition of a set of rules to calculate the
similarity between the domains of interests of the
member and each of the domains of the LO; and (2)
the proposition of a pseudo usage matrix for the
prediction of evaluations using the CF approach,
which is based, both, on the members’ evaluations
and on the members’ expertise levels and
importance degrees of the domains of the LOs.
The remainder of this paper is organized as
follows: Section 2 presents a litterature review about
recommendation systems and approaches in the
technology enhanced learning. Section 3 proposes a
personalized recommendation approach of LOs
within an online CoP. A prototype of the proposed
recommendation system and the experimental results
are presented in Section 4. Finally, the main
contribution and some future perspectives are
discussed in the conclusion.
Berkani L. and Nouali O..
Semantic Collaborative Filtering for Learning Objects Recommendation.
DOI: 10.5220/0004550500520063
In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval and the International Conference on Knowledge
Management and Information Sharing (KDIR-2013), pages 52-63
ISBN: 978-989-8565-75-4
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
2 LITTERATURE REVIEW
Recommender systems aim to generate suggestions
about new items or to predict the utility of a specific
item for a particular user.
2.1 Recommendation Approaches
Three types of approaches are distinguished: (1) the
content-based filtering (CBF) recommenders, are
built on the assumption that a person likes items
with similar features to those of other items he
preferred in the past (Peis et al., 2008); (2) the CF
recommenders, generates suggestions about data
items that users with similar tastes and preferences
liked in the past (Shafer et al., 2007); and (3) the
hybrid recommenders try to overcome the
shortcomings of the two previous approaches by
combining them in different methods (Burke, 2007).
With the advent of the semantic web, a new
generation of recommender systems based on
ontologies has emerged. These approaches take
advantage of the enhanced semantics representation.
2.2 Related Work
The state of the art shows a large number of
recommendation systems proposed in the context of
formal education, i.e. including learning offered
from educational institutions (e.g. universities,
schools) (Manouselis et al., 2009). A discussion of
the advantages and limitations of different
techniques applied in this context was presented in
(Drachsler et al., 2008). However, few works have
been proposed in an informal setting (Ziovas et al.,
2010).
As reported by Manouselis et al. (2009) an
informal setting is described in the literature as a
learning phase of so-called lifelong learners who are
not participating in any formal learning and are
responsible for their own learning pace and path
(Colley et al., 2002; Longworth, 2003). Online
communities and social networks are examples of
such contexts.
We mention for example the following
recommendation systems proposed in an informal
setting: (1) the QSIA system (Questions Sharing and
Interactive Assignments) to share educational
resources, evaluation and recommendation in the
context of online communities (Rafaeli et al., 2004).
(2) The ReMashed system for learners in informal
learning network (Drachsler et al., 2009). The main
objective of this system is to offer personalized
recommendations from the emerging information
space of a community.
The review of the literature shows that most
systems provide resources (Tang and Mccalla, 2003)
and / or individuals (Recker and Wiley, 2003),
which can help in a learning activity. Other systems
recommend courses, offering some advices to
learners for their registration in training sessions
(Garcia-Molina, 2008), or appropriate activities and
their execution sequences, allowing learners the
selection of the appropriate activities to achieve
some educational objectives (Hummel et al., 2007).
The lack of work in an informal education
motivated us to apply this approach in the context of
online CoPs. Our goal is to propose a personalized
recommendation approach taking into account the
advantages of existing hybrid systems, especially in
the domain of e-learning which is very close to our
context of study.
The proposition of a recommendation approach
in CoPs is necessary because existing systems in e-
learning, for example, can not be used directly in the
community. Learning is informal, participation
being unsupervised and the objectives and
constraints are different. In our case, the
personalization will take into account other
parameters linked, for example, to member's
expertise, skills, purpose, etc. In addition, the
representation of the resource will also take into
account the evaluation aspect according to several
dimensions: feedback, results, analysis, etc. We will
focus in this paper on the members’ profile, taking
into account some specific dimensions that are
important in the context of a CoP such as the
member’s objective, his interests and expertise.
3 A RECOMMENDATION
SYSTEM FOR COPS
We propose in this section a personalized
recommendation system for CoPs of teachers.
3.1 Recommendation Strategies
As illustrated in the Figure 1, three recommendation
strategies are proposed, according to the member’s
objective:
1. Strategy 1: If the objective is a “Specialization”,
then the system applies a SemF by domains of
interests.
2. Strategy 2: If the objective is a “Learning”, then if
there are enough ratings the system applies a CF,
language). Similarly, if the member prefers textual
resources, then the system will remove the
multimedia resources, etc.
The recommended resources will be assigned
with priorities taking into account different
parameters, such as: the difficulty of the resource
and the expertise degree of the member; see if the
resource has been visited or not, evaluated or not
(i.e. resources that are not yet visited have more
priority).
3.3 Collaborative Filtering
In the context of a CoP, members have different
levels of expertise. Accordingly, we consider that
the scores given to the resources based on the
evaluations of members should take into account this
difference of levels between members.
More formally, we propose to interpret the usage
matrix, V, taking into account the members’
expertise level for the evaluated LOs. The expertise
level of a member is calculated in relation to all of
his domains of expertise using the ontology DKOnto
(see Figure 2). We formulate the problem as follows:
Let M be a member and R a LO. Let M’s
domains of expertise be defined as a vector E
M
=
[E
1
, E
2
, …, E
m
]. We associate with this vector a
vector D
M
= [d
1
, d
2
, …, d
m
] (with same size as E
M
)
meaning that M has a degree of expertise d
i
in the
domain E
i
, where 0 ≤ d
i
≤ 1. (It should be noted that
‘m’ can be strictly smaller than the total number of
domains in the ontology.) Each resource has a set of
relevant domains denoted by vector


,
,…,
.
The idea behind this formalization of the
problem is that each domain of expertise E
i
of M is
similar to a certain degree to each domain
of R.
Thus, we define the similarity matrix of M’s
expertise with respect to R as an (m x
) matrix as
follows:
Similarity(M,R) =

,
…
,
⋮
,
⋮

,
…
,
(2)
where:
0 
,
1 is the similarity between M’s
domain of expertise E
i
and the domain
of R.
We know that each domain
of R has a
weight 0 ≤
≤ 1 for this resource. Each member M
has a degree of expertise 0 d
i
1 with respect to
his domain of expertise E
i
. We define the degree of
expertise of M with respect to
(domain j of R) as
follows:
Expertise(M,
)=

,


(3)
We define an overall degree of expertise of M
with respect to R as being:
Expertise(M,R)=
,


(4)
We finally calculate the interpreted usage matrix, by
considering only the evaluations of members having
an expertise degree greater than or equal to 0.5 for
R, as follows:




,
0,5
0, 
(5)
The obtained matrix will be used in the two steps of
the CF: (1) to calculate the similarity between the
members and infer communities, and (2) predict
notes for resources and select only those with a high
score. The evaluation consists to give a score (1-5),
from very bad to very good. Accordingly, we chose
the Pearson similarity correlation, for the prediction
of the evaluations.
3.4 Hybrid Filtering
We have proposed different methods combining the
semantic and the CF approaches. We present in this
paper two algorithms of hybridization as follows:
3.4.1 A Semantic Boosted CF Approach
The main idea is to apply a SemF, then provide
suggestions through a CF. The SemF is applied to
each row of the matrix and gradually generates a
pseudo matrix, PV. Each row, i, of this matrix
includes the evaluations given by the member M
i
, if
they are available; otherwise the predictions
calculated using the SemF are considered:




,



,
(6)
where:
v
ij
refers to the score given by the member M
i
on the
resource R
j
,
s
ij
refers to the score calculated using the SemF.
The system applies a semantic recommendation
and then the similarity results are converted into a
set of scores from 1 to 5, as follows:
If Similarity (M
i
, R
j
) [0, 0.2] then score =1
Elseif Similarity (M
i
, R
j
) [0.2, 0.4] then score =2
Elseif Similarity (M
i
, R
j
) [0.4, 0.6] then score =3
Elseif Similarity (M
i
, R
j
) [0.6, 0.8] then score =4
Elseif Similarity (M
i
, R
j
) [0.8, 1] then score =5
Finally, the CF is applied using the PV matrix.
3.4.2 A Feature Combination Approach
We propose an approach which combines the CF
and the SemF approaches using a distance formula.
The collaborative distance represents the correlation
between resources using the Pearson function, while
the semantic distance represents the similarity
between the resources using the "similarity rules",
we have proposed in section 3.2.3.
We adopt a combination method to enrich the
neighborhood, combining both semantic and
collaborative distances, using the following formula:
DistanceCo
l
DistanceSe
m
Distance/
2
7
where:
Col-Distance refers to the collaborative distance,
Sem-Distance refers to the semantic distance,
Distance represents the distance between the
resources.
The recommendation will be based on the value
of a predefined threshold, t. A set of resources will
be suggested to the member where the value of
“Distance” is greater then or equal to “t”.
4 RESULTS AND EVALUATION
4.1 ReCoPSyst: A Prototype
of a Recommendation System
In order to illustrate our work, we have developed a
personalized recommendation system called
“ReCoPSyst”, based on the proposed approach. In
order to evaluate this system, we considered a CoP
called CoPHEduc (CoP Higher Education), made up
of actors who are interested to teaching in computer
science in the university.
Figure 6 shows a screenshot of the proposed
recommendation system for this community. The
prototype ReCoPSyst was included in the
CoPHEduc portal. We can see the personalized
space of the member M1, offering for instance the
following functionalities:
Personalized recommendation of LOs and
members.
Last visited LOs.
Notifications about new added members, new
LOs, etc.
ReCoPSyst offers different recommendation
services based on the proposed approaches:
A Semantic recommendation service based on
the similarity measures. Furthermore, we have
developed other similarity recommendation
services using some existing metrics such as
Wu and Palmer (1994).
A collaborative recommendation services
using different similarity functions (Pearson,
cosine...) and according to different
recommendation types (user-user or item-
item).
Hybrid recommendation services using the
above mentioned algorithms (e.g. a semantic
boosted collaborative approach and a feature
combination approach).
We can see in Figure 7 an example of a
collaborative recommendation service. The member
can see the description of each recommended
resource, download or evaluate it. Furthermore, the
system proposes additional information about the
evaluations made by other members for each
resource (e.g. the average resource assessment, the
number of evaluators).
4.2 Tests and Evaluation
We present in this section the results of two
experimentations: (1) a qualitative evaluation; and
(2) an offline evaluation, using some existing
datasets.
4.2.1 Qualitative Evaluation
An experimental study was conducted to explore the
benefits of using the recommender system within
CoPHEduc. We describe in this section the results of
an investigation we have made to evaluate
ReCoPSyst prototype. Fifteen teachers from the
community were asked to use ReCoPSyst and then
each one provided us with a detailed feedback of
use. We have gathered more than 350 resources
from different websites such as Amazon.
Furthermore more than 300 resources were captured
by members using the system. The resources are
related to some domains of our DKOnto. The
distribution of the domains of relevance of resources
and domains of interests of members by the selected
domains is described in the table 2 below. Figure 8
illustrates this distribution, given that each resource
may be linked to several domains, and similarly,
each member may have many domains of interests.
The questionnaire of evaluation, we have
proposed, includes ten questions using a five-point
Likert scale (SA, strongly agree; A, agree; U,
undecided; D, disagree; SD, strongly disagree). The
questions are classified under four dimensions: (1)
usability, in terms of facility of use and quality of
presentation; (2) effectiveness, in terms of pertinence
such as Amazon, Book Crowsing and Merlot. The
main objective by this evaluation is to identify
which strategy is more suitable in the context of a
community of practice of teachers. Furthermore, it
will be necessary to validate the recommender
system comparing its effectiveness with other
systems on the topic such as the recommender
system for CiteSeer (Kodakateri et al., 2009) or the
recommender system proposed by Cobos et al.
(2013).
In addition, in order to improve the response
time of the proposed recommendation services, it
will be interesting to enrich our approaches using the
classification techniques.
REFERENCES
Berkani, L., Chikh, A., 2009. Towards an Ontology for
Supporting Communities of Practice of E-learning
"CoPEs": A Conceptual Model. In U. Cress, V.
Dimitrova, and M. Specht (Eds.) EC-TEL 2009. LNCS
5794, pp. 664–669. Springer, Heidelberg,
Berkani, L. and Chikh, A., 2010. A Process for
Knowledge Reuse in Communities of Practice of E-
learning. In Social & Behavioural Sciences Journal
(Elsevier) – WCES2010, 4-8 February, Istanbul,
Turkey.
Berkani, L., Chikh, A. 2013. A contextual semantic
representation of learning assets in online communities
of practice. International Journal of Metadata,
Semantics and Ontologies (IJMSO), Vol. 8, No. 1, p.
28-45, InderScience Publisher.
Burke, R., 2007. Hybrid web recommender systems. P.
Brusilovsky, A. Kobsa, . W. Nejdl, (Eds), The Adaptive
Web, vol. 4321 LNCS, chapter 12, p. 377–408.
Springer
Cobos, C., Rodriguez, O ,Rivera, J. Betancourt, J.,
Mendoza, M., León, E. Herrera-Viedma, E., 2013. A
hybrid system of pedagogical pattern
recommendations based on singular value
decomposition and variable data attributes.
Information Processing & Management 49, p. 607–
625.
Colley, H., Hodkinson, P., & Malcolm, J. 2002. Non-
formal learning: mapping the conceptual terrain.
http://www.infed.org/archives/e-texts/
colley_informal_learning.htm.
Drachsler, H., Hummel, H., Koper, R., 2008. Personal
recommender systems for learners in lifelong learning
networks: the requirements, techniques and model. In
Int. J. Learning Technoogy, vol. 3, n° 4, p. 404–423.
Drachsler, H., Pecceu, D.: Arts, T., Hutten, E., Rutledge,
L., Van Rosmalen, P., Hummel, H.G.K., Koper, R.,
2009. ReMashed - Recommendations for Mash-Up
Personal Learning Environments. In Cress, U.,
Dimitrova, V., Specht, M. (eds.), EC-TEL 2009. vol.
5794 - LNCS, p. 788-793, Springer.
Evangelou, C-E., Karousos, N., Tzagarakis, M., Gkotsis,
G., Nousia, D., 2006. Personalization Services in
Argumentation Tools: a Catalyst for Learning. In TEL-
CoPs’06, 1st International Workshop on Building
Technology Enhanced Learning solutions for CoPs,
pp. 89-99.
Garcia-Molina., H., 2008. Flexible Recommendations in
CourseRank. In R. Meersman and Z. Tari (Eds), On
the Move to Meaningful Internet Systems: OTM 2008,
vol. 5331 LNCS, p. 7–7. Springer.
Hummel, H.G.K., van den Berg, B., Berlanga, A.J.,
Drachsler, H., Janssen, J., Nadolski, R. Koper, R.,
2007. Combining social-based and information-based
approaches for personalized recommendation on
sequencing learning activities, In Int. J. Learning
Technology, vol. 3, no. 2, p.152–168.
IMS LIP, 2001. IMS Learner Information Package
Specification. The Global Learning Consortium.
Available on line at: http://www.imsglobal.org/
profiles/index.html.
Kodakateri, A., Susan Gauch, P., Luong, H.P, Eno, J.
2009. Conceptual recommender system for CiteSeerX.
RecSys 2009, p. 241-244.
Longworth, N. 2003. Lifelong learning in action -
Transforming education in the 21st century. Kogan
Page, London.
Manouselis, N., Drachsler, H., Vuorikari, R., Hummel, H.,
Koper, R., 2009. Recommender Systems in Technology
Enhanced Learning
, RecSysBook_TEL.
PAPI, 2000. Draft Standard for Learning Technology -
Public and Private Information (PAPI) for Learners
(PAPI Learner). IEEE P1484.2/D7, 2000-11-28.
Available on-line: http://edutool.com/papi.
Peis E., Morales-del-Castillo, J. M., Delgado-López J. A.,
2008. Semantic Recommender Systems. In Analysis of
the state of the topic, Hipertext.net, N 6.
Rafaeli, S., Barak, M., Dan-Gur, Y., Toch, E., 2004.
QSIA: a web-based environment for learning,
assessing and knowledge sharing in communities. In
Computers & Education, vol. 43, n°3, p. 273-289.
Recker, M., Wiley, D., 2003. A non-authoritative
educational metadata ontology for filtering and
recommending learning objects. In J. Interactive
Learning Environments, vol.9, no3.
Shafer, J.B., Frankowski, D., Herlocker, J., Sen, S., 2007.
Collaborative filtering recommender systems. In
Brusilovsky, P., Kobsa, A., Neidl, W. (eds.) The
Adaptive Web: Methods and strategies of Web
Personalization. LNCS, Vol. 4321, 291-324.
Tang, T., Mccalla, G., 2003. Smart recommendation for an
evolving e-learning system. In Workshop on
Technologies for Electronic Documents for Supporting
Learning, International Conference on Artificial
Intelligence in Education (AIED).
Tzikopoulos A., Manouselis N., and Vuorikari R., 2007.
An Overview of Learning Object Repositories. In P.
Northrup (Ed.) Learning Objects for Instruction:
Design and Evaluation, Hershey, PA: Idea Group
Publishing, pp. 29-55.
Wu, Z. and Palmer, M., 1994. Verb semantics and lexical
selection, Proceedings of the 32nd Annual Meeting of the
Associations for Computational Linguistics, Las
Cruces, New Mexico, p.133–138.
Ziovas, S., Grigoriadou, M., Samarakou, M., 2010.
Supporting learning in online communities with social
software: An overview of community driven
technologies. Advanced Technologies, InTech, p. 245-
258.
Berkani, L., Driff, L.N., Guessoum, A. 2013. Social
Validation of Learning Objects in Online
Communities of Practice Using Semantic and Machine
Learning Techniques. Modeling Approaches and
Algorithms for Advanced Computer Applications, pp.
237-247