Volume 8, Issue 7, July – 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Ontology-Based Modeling of the Learner in a

Web Educational System: Towards Learning
Analytics and Adaptive Learning
Ontology-Based Approach for Adaptive Learning

Géraud FOKOU PELAP1 Louis FIPPO FITIME2

Department of Mathematics and Computer Science Department of Computer Science
University of Dschang The National Advanced School of Engineering
Dschang, Cameroon Yaoundé, Cameroon

Bery Leouro MBAIOSSOUM3

Department of Computer Science
University of N’Djamena
N’Djamena, Chad

Abstract:- Web based e-Education systems are an activities between different platforms, between several
important type of information systems that have agents/services of a same platform, and thus achieving
benefited from Web standards for implementation, interoperability between them.
deployment and integration. One of the challenges of
these information systems is to personalize and adapt The key role of learner profile is leading
learning process according to the learners. To reach this personalization of learning experience and then to fit
goal we have to capture and model information about learning process to learner based on his knowledge and skill.
the learner, his pedagogical preferences and his learning Thus, learning activities are created and/or enriched based on
activity. In this paper, we propose an ontology-based recommendations computed from learner profiles. In
learner model, based on Semantic Web standards. We addition, learning activities are made from a set of
add this model to our models of pedagogical resources pedagogical resources in order to acquire a given set of
and, knowledge and skills model and we integrated it in knowledge and/or skills called pedagogical objectives of
an industrial context. We show how this global model these learning activities. This clearly shows an
meets industrial requirements in terms of functionalities, interdependence between knowledge/skills, pedagogical
opening up new possibilities for learning analytics and resources, learner profiles and learning activities.
adaptive learning. Consequently, description of learner profiles and learning
activities will use or reuse description of knowledge/skills
Keywords:- e-Education Information System, e-Education and pedagogical resources.
Model, Ontology, Semantic Web, Learner Profile Model,
Automatic Personalization, Adaptive Learning. Our proposed learner profile model answers the
following questions: (1) Which knowledge and skills have
I. INTRODUCTION been acquired by a learner? (2) Which pedagogical resources
have been used by a learner? (3) Which scores have been
e-Education research field has a wide range of achieved by a learner? (4) What is the history of learning
applications including : educational processes management activity for the acquisition of a knowledge/skill by a learner?
(e.g. Moodle platform), virtual classrooms, pedagogical (5) Which learning path fit learner in order to acquire a given
resources (courses, exercises, etc.), regulations (e.g. official knowledge/skill? (6) Which pedagogical resources fit to
reference standards), users management (learners and learner capabilities? We also demonstrate the feasibility of
teachers) and integration across different systems and actors our solution in a real industrial context. We integrate our
in particular to ensure compatibility and seamless user learner profile model in the Educlever's e-Education system,
experience [29]. e-Education systems leverage state of the art and we observe that the Semantic Web-based solution meets
results of Information Sciences and Technologies (IST) as industrial requirements, in terms of features, and allows
well as the Web architecture and resources to support them. Educlever System to address more requirements than its
In this context, capabilities to manage and make available existing system. Moreover, our ontology-based modelling
through the Web the description of learners as learner profile opens up new opportunities for advanced features like
as well as the description of learners activities are keys issues adaptive learning or learning trace modelling.
of e-Education research field. These descriptions meet
capabilities of sharing and/or reusing learner profiles and

IJISRT23JUL640 www.ijisrt.com 1499

Volume 8, Issue 7, July – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
This paper is organized as follows: Section II presents pedagogical resources description called ScoLOMFR [6],
state-of-the-art Educational ontologies and learners profiles based on the IEEE standard Learning Object Metadata
modeling. In section III, we recall our earlier proposed (LOM) [7] and its French version, LOMFR2. As a result, any
model which meets public standards. Section IV presents our e-Learning environment developed by public institutions or
Semantic Web based learner profile modeling for e- private companies must meet these standards and norms to
Education system. Section V shows Semantic Web based ensure a wide dissemination in e-Education context. This is
integration of our learner profile model in existing e- precisely one of the purposes of our project: develop
Educlever solution and describes implementation of solutions based on Semantic Web technologies and
Educlever features based on this learner profile model. In compatible with the standards defined by public authorities
Section VI, we propose a solution to implement adaptive and in the context of e-Education. We proposed a contribution
personalization learning on top of our propose model and for this goal in [1].
infrastructure. Section VII summarizes our contributions and
provides several perspectives. B. Ontology based User Modelling in e-Learning Systems
The learner profiles from e-Learning platforms have
II. RELATED WORK historically been the descendants of the user profiles of e-
Commerce platforms. For a long time, e-Learning platforms
A. Ontologies in e-Education represented learner profile as a user preferences model in
The interest of ontologies in the domain of e-Education order to recommend and sell online training courses as it is
has been repeatedly pointed out during the last decade. In the case in e-Commerce [28]. These preferences are also
[8], authors analyse reasons and ways to use ontologies in e- used to personalize user requests [13], and/or recommend
Education and their goals. One of these goals is pertinent answers.
implementation in e-Learning platform of features like:
management of learning institutions and/or learning
platforms and their actors, management of curricula,
management of pedagogical resources, management of
learning process and assessments [2] .

[9] presented a review and overview of works on

ontologies in the domain of e-Education and map existing
works to needs that ontologies can address. [9] classify
ontologies in e-Learning into four categories: (1) curriculum
modelling and management, (2) learning domains, (3)
learner data and (4) e-Learning services. Our proposition
could be classify in learner data. All these aspects have been
addressed by different authors. [10] propose an e-Learning
management system based on an ontology and [11] propose
ontologies built from French official texts describing
curriculum and populate them. Other ontology models like
Fig 1 Generic Learner Model.[14]
CURONTO [12] are dedicated to curriculum management
and to facilitate program review and management.
In e-learning context, [14] proposes an ontology-based
But, to the best of our knowledge, none of the model to represent generic user profiles and associated
ontologies reported in the literature has been used in an knowledge, depicted in figure 1. The main goals of this
industrial context, or evaluated on the data of an EdTech model are to share information on different users, to provide
a common vocabulary used and also to describe formally
company. Moreover, this state of the art works do not
user preferences. The learner profile contains contact
integrate public authorities recommendations or standards
model, even if there are institutional norms and standards information, relation information, preference information,
defined by public authorities or standardization committees. goal information, portfolio information. But, there is many
We point out that we speak of e-Education when e-Learning features in e-Education systems that a learner model based
solely on preferences couldn't help to reach. Among these
is applied in the public institutional context (public school or
academic) or when it has to respect the recommendations features we get, learning model, which could be define as the
ability to structure the learning path according to the
and standards of the Ministry of Public Education. In French
education, as part of the Education Code [4], the Ministry of individual skill profile of each learner [15]. We also get the
Education defined and published a common reference base assessing Learner model [15] which defines assessments,
of knowledge and skill1. It standardizes the content of with/without misconception, corresponding to a given
learner for validating a given skills or knowledge. Learner
courses by specifying knowledge and skill that a student has
model should also allow to knows all knowledge and skills
to acquire at each step of school curriculum. Moreover, the
French Ministry of Education specifies a format for digital validated by a leaner and assist him in the remediation
process [16] for those he fails to validate. Another important
feature is helping leaner to keep in mind what he has learned,
1
original name: Socle commun de connaissance, de
2
compétences et de culture http://www.lom-fr.fr

IJISRT23JUL640 www.ijisrt.com 1500

 Volume 8, Issue 7, July – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
by implement the forgetting curve [17] and determine when the e-Education solution. In the Educlever system, an
learner need to revisit some knowledge and skills. At last, example of Cocon could be the multiplication of two
considering that learners do not have the expertise to define integers identified with URI refeduclever: Multiply Two
their learning content and that Educlever's target is public Integers3.
educational institutions, the proposed learner model have to
allow learning process to meet public standards and Figure 2 shows Referential ontology model
recommendations and this could not be possible when using implements in Educlever system. This figure shows that we
solely preferences. formalize the concept Cocon as an equivalent class to EKS
from the ontology Eduprogression [11]. Thanks to
In the state of the art of learner profile design in e- formalization, we extend Eduprogression and then to use
Learning context, we find works which mainly modelling properties of concept EKS from ontology Eduprogression
learner's preferences. [18] proposed a framework to model (has Course, has Cycle and has Learning Domain) to
learner's preferences and recommend learning content based describe a Cocon.
on collaborative and content based approach. In a similar
approach, [2] and [19] propose a learning recommendation
based on a learning profile represented by a fuzzy tree. In
this recommendation system, learner model includes the
learner’s background, learning goal, required learning
categories, and learning activities are used in the
recommendation process. They used fuzzy logics to handle
similarities between concepts and applied fuzzy tree-
structured data model to model the learner profiles. [20]
propose an ontology based learner profile modeling which
describes the following concepts: knowledge, skills,
preferences or habits and interaction. They show the way an
educational resource is assigned according to the learner's
preference states during learning activities.

Unfortunately, to the best of our knowledge, existing

works do not implement important features described above
like monitoring learning path, helping learner to achieve Fig 2 Referential Ontology
institutional pedagogical objectives and keep in mind what
he has learned and recommend him/her a fit pedagogical For instance, multiplication of two integers is a
resource. Moreover, we did not find any integration and Cocon of cycle two (has Cycle), its learning domain is the
evaluation of e-Education system in a real industrial context. first domain of French education standards, languages for
We propose to overcome these limits with Semantic Web thinking and communicate (has Learning Domain) and its
models and technologies to design a learner profile that course is Mathematics (has Course). So, with this
makes it possible to find the history of learner activities. We formalization, we integrate public standards description, and
integrated this work in the Educlever company industrial we improve semantic of Cocon by adding new properties,
context and evaluated it. between them, (i) skos:broader, for hierarchical relationships
between Cocon, (ii) is PrerequisiteOf, for dependency
III. ONTOLOGY MODELS OF KNOWLEDGE, relationship between Cocon, (iii) is Followed By, for
SKILLS AND PEDAGOGICAL RESOURCES chronological dependency between Cocon, (iv) is
Complexification Of and (v) is Under Standing Lever Of [1]
We had proposed two ontology-based models to which are specialization of the relation is Related To.
describe knowledge and skills referential and pedagogical
resources [1]. These models have been integrated in the B. Pedagogical Resources Modelling
Educlever software infrastructure. These ontologies reuse We describe pedagogical resources available through
and extend EduProgression ontology [11] which is the e-Education solution in the ontology Corpus, which uses
modelling the official common base of knowledge and skill. a specific vocabulary. Figure 3 describes Corpus ontology
Since our learner profile model reuse these two models, we and we could observe that the OPD class is the keystone of
briefly review them. the ontology. OPD stands for Objet Pédagogique in French
(Pedagogical Object) and it represents pedagogical resource
A. Knowledge and Skills Modelling created to learn and acquire knowledge or skills.
We describe knowledge and skills in the ontology
called Referential which contains all the elements of
knowledge and skill available through the e-Education
solution. The concept Cocon, which stands for compétences
et connaissances in French (knowledge and skills), is the
keystone of Referential ontology. a Cocon represents an
3
atomic element of knowledge or skills learnt by students on refeduclever:
http://www.educlever.fr/edumics/refeduclever\#

IJISRT23JUL640 www.ijisrt.com 1501

Volume 8, Issue 7, July – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
have property foaf:topic_interest and then we are able to
manage learner preferences like existing works.

Fig 3 Corpus ontology

There are two key properties, properties works On and Fig 4 Learner Profile Model
has OPD. Property works On relies an instance of OPD to an
instance of Cocon, from the Referential ontology, tackled in
the pedagogical resource. worksOn is specialized into is
Learning Of, is Trainning Of, and is Evaluation Of
properties. These properties describe the role of an OPD in
the learning process of a Cocon. Property has OPD links two
OPDs. It represents partonomies and expressing how some
pedagogical resources are built as a combination of other
pedagogical resources.

Autonomous OPD is a subclass of OPD gathering the

resources which do not need any other resources to be used.
Thanks to Corpus model, e-Education company could
provide pedagogical resources annotated on public standards,
class Course and Learning Domain from Eduprogression.
Moreover, private companies could share pedagogical
resources when theses pedagogical resources allow to learn Fig 5 Learner Description with FOAF Ontology
or evaluate many different skills and knowledge.
But here, preference will be defined as union of Cocon
IV. ONTOLOGY BASED LEARNER MODEL and OPD, to show the knowledge and skills appreciated or
wanted by the learner as well as the pedagogical resources.
This section presents our proposition of learner profile Our model describes learners group and allow them to share
model based on Semantic Web technologies. In addition to interested Cocon and OPD in these groups. A deep analysis
the explicit description of a learner, our model allows an of learners activities will also allow us to recommend learner
integration of public standards and recommendations and group to learner based on their activities and also their
effective implementation of Educlever use cases involving interested topic. We recall that interested topic does not only
learner profile. More precisely, our model describes means a topic like by learner but it also could be a topic that
relationships between learners and their preferences. It also a learner have to acquire or mastered in his curriculum. The
allows the description of the learning progress over time and ontology depicted in figure 4 allows use to build a learner
allows evaluation of this learning process. social network, on top of the FOAF ontology, where learner
could share knowledge, pedagogical resource solutions and
A. Learner Profile Modelling learning process experience. So, the next step is the
Our proposed learner profile model describes learner description of the learning process.
concept. To implement this intrinsic description of learners,
we reuse the concepts of the Friend-Of-A-Friend (FOAF) B. Learner Activities Modelling
ontology [21]. Indeed, as depicted in figure 5, we observe Learning process of a learner is defined by the set of
that the class Learner inherits from the class User which activities perform by him. Then, our second contribution is
itself inherits from the class Person of the FOAF ontology. the description of these learner activities in our learners
We designed the class User because learner is not the only profile model. In order to describe learner activities, we
kind of user of e-Education platform, there are others users designed the model depicted in figure 5. A key concept of
like teachers. Since Learner is a subclass of User, Learner this model is the concept Profile cocon which represent
acquisition of a Cocon, mentioned using property has Cocon,

IJISRT23JUL640 www.ijisrt.com 1502

 Volume 8, Issue 7, July – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
through a set of OPD evaluation. The concept Profile learner_1 who learns Identify Base Sentence Components,
represents the profile of a learner, it is a set of Profile_cocon which is a Cocon.
that learner begins acquisition or already acquires. The
concept Profile contains learning path of a learner for a given In order to acquire this knowledge and skill, a learning
Cocon and for a given period of time. Indeed, learning path path is Identify Base Sentence Components T.id-ent. CI.
for a given Cocon is the set of Cocon that learner have to ident. sujet Identify Base Sentence ComponentsT. ident.
acquire and also activities on these Cocon. This set of Cocon CIIdentify Base Sentence Components, where they are
are Cocon related to the given Cocon with properties is related with property is Parent Of.
Parent Of, is Prerequisite and so on, from Referential
ontology. A curriculum of a learner contains many Cocon Figure 6 also shows that profile_learner_1 of
and all of them do not have the same pertinence in the learner_1 has two Profile_cocon (the way the user learns a
curriculum. The property has Weight of Profile_cocon Cocon): profile_cocon_1, for Cocon Identify Base Sentence
represent the weight of the Cocon in the curriculum. Components, and profile_cocon_2 for Cocon Identify Base
Similarly, acquisition of a Profile_cocon need many Sentence ComponentsT. ident. CI. Each of these
evaluation represented by the concept Evaluation_measure. Profile_cocon have two Evaluation_measure which
Since, all Evaluation_measure do not have the same represent assessments performed online by learner_1 on
pertinence for the acquisition of the given Cocon, each different pedagogical resources in different day with their
Evaluation_measure have a weight denoted by the property score. A Profile_cocon could have more than two
has Weight. Evaluation_measure, or less than two. These scores of an
assessment are over 100 marks and allow Educlever system
Now, we consider the example depicted in figure 6. to compute average using weight and then validate or not
This figure 6 shows a part of learning activities of learner acquisition of the learning path.

Fig 6 Learner Profile Instances Example

V. INTEGRATION IN SEMANTIC WEB BASED architectures were built on top of triple stores to process
EDUCATIONAL SYSTEM RDF data from the Referential and Corpus datasets.

Integration of our model takes place at two levels: (i) at In simple architecture, we load learner profile
the level of the architecture of the Educlever platform and ontology and instances in a single triplestore as depicted in
(ii) at the functional level. Figure 7. Thanks to the flexibility of the Semantic Web
solution and precisely the use of URIs to identify resources,
A. Knowledge Extraction and Integration this update does not affect existing features already
After modelling the learning profile, we have faced the implemented. Thus, SPARQL queries implementing these
challenge of deploying this model in Educlever context and features return exactly the same results. All that remains is to
migrating existing data, stored in relational databases. We implement the functionalities relating to learner profiles. But
upgrade our proposed architecture [1]: (i) Simple with this architecture, in case of failure of the triple store
Architecture and (ii) Federated. We remind that these data will no longer be available. In industrial context this is
an important risk which we have to prevent.

IJISRT23JUL640 www.ijisrt.com 1503

Volume 8, Issue 7, July – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
B. Knowledge Engineering: Learning Analytics
Now, we are able to address Educlever's existing needs
and the features they wish to add to their system. Among
Educlever's existing needs dealing with learner profile, we
find: R1 The search for administrative information on a
learner, R2 The number of evaluations carried out by a
learner, R3 . The score of the evaluations carried out by the
learner, thus this uses case allows to know if the evaluation
was passed successfully or not, R4 . The most recent
evaluations carried out by the learner and used to measure
his acquisition of the knowledge and skill assessed (Cocon).

In addition to the above listed use cases implemented

in the Educlever platform, using Semantic Web technologies
allows to implement additional features. Among these new
Fig 7 Simple Architecture features we have: R5 find information on learning resources
that have been used to assess a learner, this feature avoids
In the federated architecture, we add to our federation a redundancies in a given time interval and thus avoids
third triple store for learner profile ontology and instances as distorting the assessment. R6 Retrieve all the knowledge and
depicted in Figure 8. Thus, the federated endpoint allows to skill (Cocon) acquired by the learner. R7 Determining
query, in a transparent way, the three datasets. Thanks to whether a learner has all the prerequisites to start learning of
Semantic Web technologies, this operation is made with a given Cocon, it is a direct consequence of the previous
configuration instructions in an industrial production context. feature. R8 Our model also allows us to assist a learner in the
Moreover, this architecture prevents failure of one of the remediation process [16] in the event of failure on a given
triple store, and allows to query others as a single dataset. Cocon learning. For this, we are looking for Cocon already
This context and scenario is typical of the need to take into acquired by the learner or for which he has all the
account legacy software and information system from real prerequisites and which are levers for understanding (is
industrial contexts as well as the service quality constraints, Under Standing Lever Of) of the Cocon to acquire. R9 . We
etc. have integrated an implementation of the forgetting curve
[17] in order to accurately get acquisition level of a Cocon
over time

Implementation of these features has been done

through one SPARQL query or functions executing several
SPARQL queries. While existing features have been
implemented in Educlever platform with only functions and
execute many queries. This shows useful of Semantic Web
based platform since it implement more features than the
existing Educlever platform. Table 1 compares existing
Educlever Platform and Semantic Web platform
effectiveness in uses cases implementation. This table shows
use case implemented on the existing Educlever platform
and our proposed upgrade.

Fig 8 Federated Architecture

Table 1 Implementation of the use Cases

Existing Features Added Features
R1 R2 R3 R4 R5 R6 R7 R8 R9
Existing Educlever Platform    
Educlever Semantic Web Based Platform         

In the previous table 1, case with symbol  means that Then, we could observe that using Semantic Web Based
we implement these uses cases with one query. While platform we meet all the uses cases. The new platform,
symbol  means the implementation use a function which could also lead the personalization of learning experience
itself executes several queries in order to fill the use case. In [25, 26]. For a best learning experience, predicting learner’s
table 1, empty cells means that use case is not implemented. success on pedagogical resources is an important step.

IJISRT23JUL640 www.ijisrt.com 1504

 Volume 8, Issue 7, July – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
VI. ADAPTIVE LEARNING IN SEMANTIC WEB Factorization machine (FM) is a machine learning
BASED EDUCATIONAL SYSTEM algorithm based on matrix decomposition proposed by
Steffen render. The form of Factorization machine is like:
A. Predicting Learner Success based on the Learning
Resources Knowledge Graph
Let U={u1, u2,…, u|U|}, I={i1, i2,…, i|I|} and S={s1,
s2,…, s|K|} represent the sets of learners , items (OPD) and
Skills (Cocon) respectively. The task of this section focus on
predict the outcome of learner towards a specific item, and
then recommend learner items that he/she will need to Among them vj,f and vi,f are respectively a hidden
improve his/her master of a certain knowledge. Each learner factor of the corresponding hidden vector of character i and j.
is associated with a sequence of items from the items set . Usually, because of the sparse data, we cannot learn w, but
Each learner will try a single item for a several times. By we can learn the parameter vector v of and features
sorting interaction records in a chronological order, we can respectively through the data of features and other features.
form the interaction sequence for user So we predict the value of w through the product xj*xi,
which solves the data sparsity problem.
u. The index denotes the relative time index, instead of
absolute time index in temporal recommendation. With the
B. Experimentation and Evaluations
above notations, given learner u history sequence , we aim In this section, we implement the prediction model and
to infer the item(s) that a learner is likely to do which will evaluate its capacity to predict success on learner training.
help him/her quickly learn a certain knowledge.
 Datasets
To perform this task we going to setup a knowledge We used two dataset in these experiments: (i)
graph embedding model. The knowledge graph embedding Assistmens09: The dataset of Assessments [24] described
algorithm transforms the knowledge graph into a low- in (Feng, Heffernan and Koedinger 2009) which is one of
dimensional dense real number vector which is used as input the dataset used in state of the art work KTM with 4217
of existing machine learning algorithms: the vector students over 26688 questions, 123 knowledge components
representation of nodes can be sent as a feature to support (OPD) and 347k interactions. (ii) Educlever knowledge
vector machine and other classifiers for node classification graph 1330 users attempting 334k items, 17127 skills with
and prediction tasks. We used the knowledge tracing to 85270 times interactions.
encode training records. The most popular model is Bayesian
knowledge tracing (BKT), which is a hidden Markov model  Data and Encoding of Interaction
[27]. It is used in many intelligent tutoring systems to model Here, we present how to encode the observed data into
each learner's mastery of the knowledge being tutored and sparse vector x. First, we need to choose features which will
models student knowledge in a Hidden Markov Model as a be used in modeling :
latent variable, updated by observing the correctness of each
student's interaction in which they apply the skill in question.  Users: Assume there are n learners, the first n feature
But for this work, we choose to use Knowledge Tracing will be for n learners, and in KTM, it use one-hot
Machines (KTM). encoding. Let’s say if learner i is involved in this
interaction, then its xi value will be 1, the rest for the
Knowledge Tracing Machines (KTM) [22,23] is a other learners set to 0.
sequence prediction problem where the goal is to predict the  Items: Assume there are m items (OPD). So there will
outcomes of students over questions as they are interacting be m features to represent m items. If item j is involved
with a learning platform. By tracking the evolution of the in one interaction, then are setting to 1, the rest items
knowledge of some student, one can optimize instruction. It
remain 0.
uses factorization machines (FMs), a model for regression
or classification, encompasses several existing models in the  Skills: Let us assume there are s skills (Cocon). There
educational literature as special cases to estimate student will be s features to represent s skills. If one interaction
knowledge accurately and fast even when student data is involves several skills, the corresponding index of skills
sparsely observed, and handle side information such as are setting to 1.
multiple knowledge components and number of attempts at  Wins & Fails Allocate s features to distinguish if a
item or skill level. This approach allows to fit student models user learn a skill is success if the attempt was correct, s
of higher dimension than existing models, and provides a more features as opportunities to have learned a skill if
testbed to try new combinations of features in order to attempt was incorrect.
improve existing models. In this work, we use KTM as state
of the art algorithm, to predict learner’s behavior towards Table 2 is an example for encoding of users, items,
OPD learner has never interacted before, and based on the skills, wins, and fails. Here, we have n=2 learners, m=3
outcome, to recommend learner OPDs to improve his/her questions, s=3 skills. The first row is an interaction of
mastery of a certain Cocon. learner 2 tried question 2, and question 2 involves skill 1
and skill 2. At the beginning, learner 2 has no interaction
with any question before, so he/she doesn’t have the chance
to learn any skill, so the count of wins & fails for any skill

IJISRT23JUL640 www.ijisrt.com 1505

Volume 8, Issue 7, July – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
are all 0. Then we can see the outcome is 1 for question 2,
so in the second row, the wins for skill 1 and skill 2 add 1.
Therefore, we encode triplets with N = n + m +3s = 14
features.

Table 2 Example of Encoding for Training in KTM  Metrics and Evaluation

To evaluation our model we going to use two
standards metrics:

 ACC: Stand for ACCuracy. The accuracy rate

represents the proportion of correctly classified samples
to the total number of samples. If we predict 10 samples
and 8 samples are correct, the accuracy rate is 80%.
 AUC: Stands for Area Under the Curve. It is defined as
the area below the ROC curve: The horizontal axis of
ROC curve is true positive rate (TPR), the vertical axis
is false positive rate (FPR), also known as true positive
rate. AUC can only be used for the evaluation of two
classification model.
 Extract Data from Educlever Knowledge Graph.  NLL: Stands for Negative Log-Likelihood. This metric
We use the SPARQL query to extract user interactions becomes unhappy at smaller values, where it can reach
from the Educlever Knowledge Graph, more specific, from infinite unhappiness, and becomes happy at larger
profile graph. In order to be able to use Educlever dataset in values. Because it is summing the loss function to all
the state of the art work KTM, we get user, item, skills the correct classes, what’s actually happening is that
from the graph, item and skill corresponding to OPD and whenever the network assigns high confidence at the
Cocon in Educlever Knowledge Graph. And the interactions correct class, the unhappiness is low, but when the
should be grouped based on user and item in a chronological network assigns low confidence at the correct class, the
order. So, we use the following SPARQL query to get the unhappiness is high.
data:
For our evaluation, we run experiments for 5 times and
average the result. We first run the experiment on dataset
Assessments09. When encode interactions only with user
and item, got the result:

Table 3 Experiment Results on Assessment09

Model ACC AUC NLL
Users, Items 0,726 0,730 0,571
Users, Items, Skills 0,703 0,709 0,580
Users, Items, Skills, Win, Fail 0,737 0,730 0,545

For the experiment on Educlever Knowledge Graph,

the results are in the table below:

Table 4 Experiments Results on Educlever KG

Note that in the encoding of KTM, we use, correct = 1
Model ACC AUC NLL
or 0 to denote learner’s performance towards item. So, we
Users, Items 0,833 0,695 0,424
convert the score from Educlever to 0 or 1 to match the
algorithm. Users, Items, Skills 0,832 0,80 0,383
Users, Items, Skills, Win, Fail 0,855 0,906 0,280
 Data Processing
Note that data stored in Educlever Knowledge Graph We observe that using KTM to make prediction on
are mostly URIs, but according to KTM, every learner, Educlever Knowledge Graph more relevant and get satisfied
every item and skill should have a unique ID which could be results with 0.885 accuracy and 0.906 AUC when encoded
used to do one-hot encoding. So, after got data, the first step with wins and fails. And, from the two experimentation, we
is to process the data to match the algorithm. For learner, the observe that the last model, where all the feature are
URI is like http://www.educlever.fr/edumics/ refeduclever# encoded, is the more efficient for the prediction.
USER_10. The last part of URI is number, only extract
number will be fine. The same works for items. However, in After running the experiment, we will get a result file
Educlever Knowledge Graph knowledge graph, there is no which consist of probabilities that a user will do it correctly
id for skills. Therefore, we perform a query to get all skills towards the item. Based on these probabilities, we sort these
in the graph, and then randomly mapping allocate them probabilities, and select the top 10 items that user tend to do
unique id. badly, and recommend user correspondent skills to practice.

IJISRT23JUL640 www.ijisrt.com 1506

Volume 8, Issue 7, July – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Because user will try the same item for several times, ACKNOWLEDGMENT
so there has more than one records for a user towards a
single item. Thus, during the prediction, we will get Thanks to Educlever4 to be part of this project and
different probabilities for the same user-item. Therefore, the share data with us. Thanks to WIMMICS/INRIA5 to hire me
first step is to keep only one record with the highest for this project. And finally, thanks to the company
probability for a single user-item. And then sort all the items Modelling Innovative Intelligence Artificial (M2IA) for the
that the user has interacted with. technical and financial support.

VII. CONCLUSION REFERENCES

This work reported a knowledge modelling experience [1]. G. Fokou Pelap, C. Faron-Zucker et F. Gandon,
in industrial context to propose an e-Education solution «Semantic Models in Web based Educational System
based on Semantic Web models and technologies. We Integration,» chez 14th International Conference on
recalled our previous work on modeling knowledge and skill Web Information Systems and Technologies
(Cocons), as well as pedagogical resources (OPD). Then we (WEBIST), Seville, 2018.
presented our proposal of an ontology describing a shared [2]. K. Balaji et G. Poorni, «An Effective
conceptualization of a learner, learner activities over time Recommendation System for E-Learning Using
and learner preferences, that can smoothly extend our Fuzzy Tree,» Semantic Scholar Journal, 2017.
modelling in order to implement user-oriented use cases. [3]. M. Srivastava, H. Pandey, S. Kumar Shukla et B.
Kumar Thakur, «A Literature Review of E-Learning
We also proposed two extensions of Educlever Model Based on Semantic Web Technology,»
architectures proposed in our previous work, which allows us Semantic Scholar Journal, 2014.
to integrate and process learner profiles ontology and [4]. Nationale Ministère de L'éducation, «Code de
instances in the Educlever platform. Thanks to Semantic l'éducation consolidé au 1er Janvier 2018,» Service
Web technologies, we integrate learner profiles without any Publique, 2018.
change on features already implemented which continue to [5]. Nationale Ministère de L'éducation, «Le socle
work properly. This integration allows us to implement commun de connaissances, de compétences et de
existing features on learner profiles, as well as features culture,» Service Publique, 2017.
which could not be implemented on the existing Educlever [6]. Réseau Canopé, «ScoLomFR : Outil de description
infrastructure. To implement these features using Semantic des ressources numériques,» Service Publique, 2011.
Web technologies we use, depending on the complexity of [7]. IEEE-LTSC-LOM Committee, «IEEE Standard for
the feature, either a SPARQL query or a function executing Learning Object Metadata,» IEEE Standards
several SPARQL query. Based on these features, we are able Association, 2002.
to provide algorithms to personalize learning path for a given [8]. L. Jaffro, «Les objets de l'éducation : quelle ontologie
learner. The most immediate continuation of this work is the ?,» Revue de métaphysique et de morale, vol. 4,
experimental evaluation of our model in Educlever context. n°56, pp. 429-448, 2007.
[9]. A.-Y. Maha, G. Remya et A.-F. Auhood, «Ontologies
Next to all the previous contributions, we manage to in E-Learning: Review of the literature,»
get real data from Educlever Knowledge Graph and format International Journal of Software Engineering and its
them in order to perform prediction. In this work, we are able Applications, vol. 9, pp. 67-84, 2015.
to extract useful information to make predictions and [10]. M. Rani, K. V. Srivastava et O. P. Vyas, «An
successfully make recommendations based on existing state Ontological Learning Management System,»
of the art algorithm. Computer Applications In Engineering Education,
vol. 24, n° %15, p. 706–722, September 2016.
One of the next challenges is implementation of [11]. O. R. Rocha, C. Faron-Zucker et G. Fokou Pelap, «A
adaptive and personalized learning based on learner profiles. Formalization of the French Elementary School
Indeed, we plan to customize state of the art prediction Curricula,» chez Knowledge Engineering and
algorithm and combine it with rule-based reasoning Knowledge Management – (EKM) and Drift-an-
mechanisms in order to propose learning path and LOD, Bologna, Italy, Revised Selected Papers,
recommend relevant pedagogical resources for his/her Bologna, 2016, pp. 82--94.
training. We will also propose measures to evaluate the [12]. A.-Y. Maha, A.-F. Auhood et G. Remya,
acquisition of a given knowledge and skills (Cocons). Based «CURONTO: An Ontological Model for Curriculum
on this measure, we are going to consider creation of student Representation,» chez Proceedings of the 18th ACM
groups based on pedagogical objectives, such that they have Conference on Innovation and Technology in
complementary knowledge and skills (Cocons) or same level Computer Science Education, ACM, 2013, pp. 358--
of acquisition for a given set of knowledge and skills. 358.

4
https://www.educlever.com/
5
https://www.inria.fr/fr/wimmics

IJISRT23JUL640 www.ijisrt.com 1507

Volume 8, Issue 7, July – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
[13]. G. Koutrika, E. Pitoura et K. Stefanidis, «Preference- [26]. A. Ettorre, F. Michel et C. Faron, «Prediction Of
Based Query Personalization,» chez Advanced Query Students’ Performance In E-Learning Environments
Processing, vol. 36, Heidelberg, Springer Based On Link Prediction In A Knowledge Graph,»
International Publishing, 2013, pp. 57--81. chez Artificial Intelligence in Education. Posters and
[14]. K. Rezgui, H. Mhiri et K. Ghédira, «An Ontology- Late Breaking Results, Workshops and Tutorials,
based Profile for Learner Representation in Learning Industry and Innovation Tracks, Practitioners’ and
Networks,» International Journal of Emerging Doctoral Consortium: 23rd International Conference,
Technologies in Learning (iJET), vol. 9, n°3, pp. 16-- AIED 2022, Durham, UK, July 27–31, 2022,
25, 2014. Proceedings, Durham, Springer-Verlag, 2022, p.
[15]. M. C. Desmarais et R. S. Baker, «A Review of 432–435.
Recent Advances in Learner and Skill Modeling in [27]. A. Bordes, N. Usinier, A. Garcia-Duran, J. Weston et
Intelligent Learning Environments,» User Modeling O. Yakhnenko, «Translating Embeddings for
and User-Adapted Interaction, vol. 22, n° 1-2, pp. 9-- Modeling Multi-relational Data,» chez Advances in
38, April 2012. Neural Information Processing Systems, vol. 26,
[16]. C. Reverdy, «L’accompagnement à l’école : Curran Associates, Inc, 2013.
dispositifs et réussite des élèves,» Dossier de veille [28]. M. Dastani, N. Jacobs, C. M. Jonker et J. Treur,
de l’Institut Française de l'Education (IFE), vol. 119, «Modelling user preferences and mediating agents in
pp. 1--31, 2017. electronic commerce,» Knowledge-Based Systems,
[17]. J. S. Fisher et G. A. Radvansky, «Patterns of vol. 18, n° 7, pp. 335 - 352, 2005.
forgetting,» Journal of Memory and Language, vol. [29]. Bery L. Mbaiossoum, Atteib D. Mahamat, L.
102, pp. 130 - 141, 2018. Dionlar, «Building a Smart education in developing
[18]. K. Abhinav, V. Subramanian, A. Dubey, P. Bhat et countries: Case of Chad», International Journal of
A. Divakaruni Venkat, «LeCoRe: A Framework for Innovation and Scientific Research, vol. 39, n° 2, pp.
Modeling Learner's preference,» chez Proceedings of 170 - 179, 2018.
the 11th International Conference on Educational
Data Mining, (EDM) 2018, Buffalo, NY, USA, July
15-18, 2018, International Educational Data Mining
Society (IEDMS), 2018.
[19]. D. Wu, J. Lu et G. Zhang, «A Fuzzy Tree Matching-
Based Personalized E-Learning Recommender
System,» IEEE Transactions on Fuzzy Systems, vol.
23, n° 6, pp. 2412-2426, 2015.
[20]. A. Korchi, N. El Amrani El Idrissi et L. Oughdir,
«Modeling and Implementing Ontology for
Managing Learners’ Profiles,» International Journal
of Advanced Computer Science and Applications,
vol. 8, n° 8, pp. 144--152, 2017.
[21]. D. Brickley et L. Miller, «FOAF Vocabulary
Specification 0.99,» 2014.
[22]. C. Piech, J. Bassen, J. Huang, S. Ganguli, M. Sahami,
L. J. Guibas et J. Sohl-Dickstein, «Deep Knowledge
Tracing,» chez Advances in Neural Information
Processing Systems, vol. 28, Curran Associates, Inc.,
2015.
[23]. H. Su, X. Liu, S. Yang et X. Lu, «Deep knowledge
tracing with learning curves,» Frontiers Psychology,
vol. 14, March 2013.
[24]. E. Lavoué, B. Monterrat, M. Desmarais et S. George,
«Adaptive Gamification for Learning Environments,»
IEEE Transactions on Learning Technologies, vol.
12, n° 1, pp. 16 - 28, March 2019.
[25]. A. Ettorre, O. R. Rocha, C. Faron, F. Michel et F.
Gandon, «A Knowledge Graph Enhanced Learner
Model to Predict Outcomes to Questions in the
Medical Field,» chez Knowledge Engineering and
Knowledge Management: 22nd International
Conference, EKAW 2020, Italy, September 16–20,
2020, Proceedings, Bolzano, Springer-Verlag, 2020,
p. 237–251.

IJISRT23JUL640 www.ijisrt.com 1508