=Paper=
{{Paper
|id=Vol-2354/w4paper6
|storemode=property
|title=Optimizing Recommendation in Collaborative E-Learning by Exploring DBpedia and Association Rules
|pdfUrl=https://ceur-ws.org/Vol-2354/w4paper6.pdf
|volume=Vol-2354
|authors=Samia Beldjoudi,Hassina Seridi
|dblpUrl=https://dblp.org/rec/conf/its/BeldjoudiS18
}}
==Optimizing Recommendation in Collaborative E-Learning by Exploring DBpedia and Association Rules==
https://ceur-ws.org/Vol-2354/w4paper6.pdf
Optimizing Recommendation in Collaborative E-
Learning by Exploring DBpedia and Association Rules
Samia Beldjoudi1, Hassina Seridi2
1
Superior School of Industrial Technologies, Annaba, Algeria
1, 2
Laboratory of Electronic Document Management LabGED Badji Mokhtar University, Anna-
ba, Algeria
1
s.beldjoudi@epst-annaba.dz 2seridi@labged.net
Abstract. Social tagging activities allow users to add free annotations on re-
sources to express user interests, preferences and automatically generate folk-
sonomies. This paper demonstrates how structured content available through
DBpedia can be leveraged to support recommendation of resources in folkso-
nomies. A limitation of resources’ recommendation is the content overspeciali-
zation conducting in the incapability to recommend relevant resources different
from the ones that the learner already knows. To address this issue, we pro-
posed to take advantage of the richness of the open and linked data graph of
DBpedia and association rules to learn learners' behavior. The proposed ap-
proach demonstrates the efficiency of using DBpedia to enhance diversity and
novelty when recommending resources to users in folksonomies. The basic idea
is to iteratively explore the RDF data graph to produce novel and diverse rele-
vant recommendations.
Keywords: Collaborative E-learning, Recommendation, DBpedia, Diversity,
Novelty.
1 Introduction
Social tagging systems have achieved a great success over the web in the last years,
especially in recommendations approaches. The problem of a precise recommender
system is that the entire set of recommended resources may be obvious as one consid-
ers the case of a film recommendation algorithm that only returns films of the same
actor. To overcome this problem, novelty and diversity should be also considered in
the evaluation of a recommender system, as precision only offers an incomplete de-
scription of the system’s effectiveness.
The main focus of our study is how to exploit the semantic aspect of DBpedia to
enhance resource recommendation within social tagging systems. We propose a new
method for analyzing learner profiles according to their tagging activities in order to
165
2
improve the recommendation of resources. The effectiveness of results depends on
the resolution of social tagging drawbacks. In our process, we demonstrate how we
can reduce the tags ambiguity problem by taking into account social similarities cal-
culated on folksonomies combined with similarities between resources in DBpedia.
We used also the force of Linked Open Data (LOD) to enhance resource recommen-
dation by exploring the interlinked entities in LOD cloud. We base up on the iterative
exploration of the DBpedia graph to obtain novel and diverse recommendations that
should satisfy the learner and create the effect of surprise by recommending resources
that the user did not expect at the beginning.
This paper is organized as follows: Section 2 is an overview of the main contributions
related to our work. Section 3 is dedicated to the presentation of our approach. In
section 4 we present and discuss the results of some experiments we conducted to
measure the performance of our approach. Conclusion and future works are described
in Section 5.
2 Related works
Social web based approaches, like folksonomies, have achieved a high level of im-
provement even in E-learning practice. In this section, an overview about some con-
tributions attached to this field is proposed. [Kopeinik et al., 2017] investigated the
application of two tag recommenders that are inspired by models of human memory.
The authors find that displaying tags from other group members helps significantly in
semantic stabilization in the group, as compared to a strategy where tags from the
students' individual vocabularies are used. In [Beldjoudi et al., 2016], the authors
proposed a new approach for personalizing and improving resources retrieval in col-
laborative learning with tackling tags ambiguity and event detection impact on re-
sourced retrieved by ranking. In another contribution [Beldjoudi et al., 2017] pro-
posed a method to analyze user profiles according to their tags in order to predict
interesting personalized resources and recommend them. The authors proposed a new
approach to reduce tag ambiguity and spelling variations in the recommendation
process by increasing the weights associated to web resources according to social
similarities. They base upon association rules for discovering interesting relationships
among a large dataset on the web. [Karabadji et al., 2018] proposed to focus mainly
on the growing of the large search space of users’ profiles and to use an evolutionary
multi-objective optimization-based recommendation system to pull up a group of
profiles that maximizes both similarity with the active user and diversity between its
members. In such manner, the recommendation system will provide high perfor-
mances in terms of both accuracy and diversity. In our work we want to leverage the
social and semantic web in order to enhance educational resources recommendation in
collaborative e-learning.
166
3
3 Approach description
In this paper, we propose a method to analyze learner profiles according to their
tags in order to predict interesting personalized resources and recommend them. We
argue that the automatic sharing of resources strengthens social links among learners
and we exploited this idea to reduce tag ambiguity in the recommendation process by
increasing the weights associated to web resources according to social similarities.
We based upon association rules that are a powerful method for discovering interest-
ing relationships among a large dataset on the web. Our goal was to find correlations
between tags, i.e. to find tags frequently appearing together, in order to extract those
which are not used by one particular learner but which are often used by other users
close to him in the social network.
The effectiveness of the recommendation depends on the resolution of the prob-
lems of folksonomies. In our approach we tackle the problems of tag ambiguity, di-
versity and novelty. To resolve the problem of tag ambiguity in recommendation, we
propose to measure the similarity between learners to identify those who have similar
preferences and therefore adapt the recommendation to learner profiles.
- First step: For each extracted association rule (Tags A → Tags B) whose antece-
dent applies to an active learner lx, we measure the similarities between this learner
and the learners of his social network who use the tags occurring in the consequent of
the rule. The resources associated to these tags are recommended to the learner de-
pending on these similarities. To measure similarity between two learners (l1 and l2),
both are represented by a binary vector representing all their tags and we compute the
cosines similarity between the two vectors.
- Second step: To avoid the cold-start problem which generally results from a lack
of data required by the system in order to make a good recommendation, when the
learner of the recommender system is not yet similar to other users, we propose to
exploit semantic links between resources in DBpedia. DBpedia can be a reliable and
rich source of content information that supports recommender systems to overcome
problems, such as the cold-start problem and limited content analysis that restrict
many of the existing systems, by building on a robust measurement of the similarities
between resources using DBpedia. In this approach, we use the Linked Open Data to
assess the similarity between folksonomies resources using their corresponding re-
sources on DBpedia (i.e. we measure the similarity between the resources that would
be recommended by the system, as related to a tag occurring in the consequent of an
association rule, and those that are already recommended to the learner). The similari-
ty between two resources is calculated using Jaccard index.
In another hand, when using a recommender system such as those of online stores,
the results are mainly expected by the users. In this case, it is clear that the recom-
mendation is not very helpful in the sense of the lack of diversity and novelty. To
solve this dilemma in folksonomies-based collaborative learning, we propose extract-
ing the most popular features found in the resources-based learner profile (i.e. the
characteristics that interest the learner when they tag their resources) and then explore
the LOD to extract resource linked with these features.
167
4
Let us consider a learner profile composed from the resources (R1, R2, R3 and R4),
Thus the intersection between the resources' features must be calculated (R1∩ R2 ∩
R3 ∩ R4), this is done because we want to extract the most popular characteristics that
interest the learner when they choose tagging their resources. Then for each feature
(Pi) in the result of intersection we will explore the LOD graph in the first level to
extract other resources (R5) having these features or having a direct/ indirect link with
these later (R6, R7 resp).
Supposing that (R1∩ R2 ∩ R3 ∩ R4) = {[domain: informatics]; [author: …]; [year:
…]; [edition: …]…}. By exploring the LOD graph we find that the resource “infor-
matics” is linked with other resources (for example: “University, Formation, Bio-
Informatics…”) via the predicates (P1, P2, P3…). In its turn the resources “Universi-
ty, Formation, Bio-Informatics…” are linked via other predicates (Pj) with other re-
sources (for example: “Boston University…”). Therefore, it appears relevant to rec-
ommend some courses of the Boston University to the current user.
Our approach is based on the iterative exploration of the DBpedia graph, where
each step depends on the result of the previous steps. In order to obtain relevant and
personalized recommendations for each learner, we calculate the occurrence number
of the {domain, author, year, edition…} characteristics and then we choose the ones
that best reflect the learner interest to exploit them later in the exploration of the RDF
graph of DBpedia.
The purpose of the graph exploration is to obtain recommendations that should not
only satisfy the learner but also to have diversity and a novelty in the recommenda-
tion, to create the effect of surprise by recommending resources that the learner did
not expect at the beginning. The learner evaluates the recommended resources in real
time in each iteration. The process stops when none of the recommended resources
has satisfied the user.
If the learner liked at least one resource among those in the proposed list, in the
second iteration, we focus on these ones. Thus, we re-explore the LOD graph again
starting from these items by using the query language SPARQL to return more educa-
tional resources connected with them; this technique allows us to propose a list of
diverse and novel resources to ensure the surprise effect.
The real-time evaluation process as well as the exploration of the graph is iterative.
At each iteration, we explore the graph based on the positive ratings assigned to the
resources previously recommended. Indeed, the evaluation is an essential step to de-
termine the new pattern of requests for the re-exploration of the graph to generate
another list of recommendations. At each step, we propose to the user 10 resources, if
he assigns a rating more or equal to three, we consider that he liked the recommended
resource, and so we record it in his profile, otherwise we move to another resource.
After evaluating the 10 resources, the program suggests to the user to recommend
after the 10 resources have been evaluated, the program suggests to recommend some
more to the user. If he accepts then another list of resources is generated from his
profile, otherwise, we stop and return the list of resources liked. With this method, we
ensure that the recommended list of resources is diverse, where every user can obtain
diverse resources even if they do not appear in the profile of his neighbors in the so-
cial network.
168
5
4 Experimental Results
In this section, experiment over a popular dataset is described and results are ana-
lyzed and discussed. The dataset exploited in our test is del.icio.us. In this experiment,
we were interested in data generated from users who tagged resources about educa-
tion. Thus, our database comprises 1128 tag assignments involving 95 users, 432 tags
containing ambiguous tags and 314 resources.
4.1 Experimental Methodology
To evaluate the quality of a recommender system, we must demonstrate that the
recommended resources are really being accepted and added by the users. Because
the knowledge of this information requires asking the users of the selected databases
if they appreciated the proposed set of resources, which is impossible in our case
because we do not have access to this community, we have used a cross-validation
where we have randomly removed some resources from the profile of each user, and
we applied our approach on the remainder dataset in order to show if it can
recommend the removed resources to their corresponding users or not. If it is the case,
so we can conclude that our approach enables to extract the user preferences.
1,6
1,4
1,2
1
Precision
0,8 Recall
0,6 F1
diversity
0,4
novelty
0,2
0
iteration 1 iteration 2 iteration 3 iteration 4 iteration 5
Fig. 1. Average precision, recall, F1, diversity and novelty of the recommendations
The curve presented in figure 1 show average values of precision, recall, F1,
diversity and novelty measures in the five iterations. We notice that the precision
achieved a good value in all iterations, this is due to the fact that the system
recommends exactly the items wanted by the user i.e. those that match his profile.
Sometimes the system begins to deteriorate in terms of precision but always with a
value that exceeds 0.6. This decrease is quite normal since the system begins to
recommend items according to different attributes (domain, year ...) which is known
as diversity of recommendation. Learners sometimes accept the recommended
resources and other times it was not the case. Recall and F1 measure achieved all both
good values in the all iterations.
169
6
To calculate individual diversity and novelty, we used the metrics proposed in
[Zhang and Hurley, 2009] and [Vargas, 2014] respectively. Figure 1 showed promis-
ing values of both diversity and novelty in the five iterations. This demonstrates the
importance of DBpedia to extract more diversified and novel resources in the recom-
mendation. It is clear that the effectiveness of recommendation depends of preserving
both precision and diversity. Results demonstrate that our approach preserving both
them in all iterations.
5 Conclusion
In this contribution we have exploited the strength of social aspect in folksonomies
to let members in the community benefit from the educational resources tagged by
other users, based on the recommendation of resources. The proposed approach is
based on DBpedia, the objective was to overcome the problem of diversity and novel-
ty in recommendation. Primary results show also the utility of exploring LOD graph
in ensuring diversity when recommending personalized educational resources in so-
cial tagging systems. In order to continue and improve our work, we aim at using
others principles like event detection, for example, to help capturing and analyzing
the behavior of learners when new events come, this can improve recommendation
and even resources ranking.
References
1. Beldjoudi, S., Seridi, H. & Benzine, A.: The Impact of Social Similarities and Event De-
tection on Ranking Retrieved Resources in Collaborative E-Learning Systems. In: Koch
F., Koster, A., Primo, T., Guttmann, C. (eds) Advances in Social Computing and Digital
Education. CARE 2016, SOCIALEDU 2016. Communications in Computer and Informa-
tion Science, vol 677. Springer, Cham (2016)
2. Beldjoudi, S., Seridi, H. & Faron-Zucker, C.: Personalizing and Improving Resource Rec-
ommendation by Analyzing Users Preferences in Social Tagging Activities. Computing
and Informatics 36(1): 223-256 (2017)
3. Karabadji, NEI., Beldjoudi, S., Seridi, H., Aridhi, S. & Dhifli, W.: Improving memory-
based user collaborative filtering with evolutionary multi-objective optimization. Expert
Syst. Appl. 98: 153-165 (2018)
4. Kopeinik, S., Lex, E., Seitlinger, P., Albert, D. & Ley, T.: Supporting collaborative learn-
ing with tag recommendations: a real-world study in an inquiry-based classroom project.
In: LAK '17 Proceedings of the Seventh International Learning Analytics & Knowledge
Conference Pages 409-418 (2017).
5. Vargas, S.: Novelty and diversity enhancement and evaluation in recommender systems
and information retrieval. In Proceedings of the 37th International ACM SIGIR Confe-
rence on Research & Development in Information Retrieval, SIGIR ’14, pages 1281–
1281, New York, NY, USA, (2014).
6. Zhang, M. and Hurley, N.: Novel item recommendation by user profile partitioning. In
2009 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent
Agent Technology, volume 1, pages 508–515, Sept (2009).
170