Building an Adaptive E-Learning System

Christos Chrysoulas

and Maria Fasli

1,2

School of Computer Science and Electronic Engineering, University of Essex, Essex, U.K.

Institute of Analytics and Data Science, University of Essex, Essex, U.K.

Keywords: E-learning, Adaptive Learning, User Profile, Learning Path, Machine Learning.

Abstract: Research in adaptive learning is mainly focused on improving learners’ learning achievements based mainly

on personalization information, such as learning style, cognitive style or learning achievement. In this paper,

an innovative adaptive learning approach is proposed based upon two main sources of personalization

information that is, learning behaviour and personal learning style. To determine the initial learning styles of

the learner, an initial assigned test is employed in our approach. In order to more precisely reflect the

learning behaviours of each learner, the interactions and learning results of each learner are thoroughly

recorded and in depth analysed, based on advanced machine learning techniques, when adjusting the subject

materials. Based on this rather innovative approach, an adaptive learning prototype system has been

developed.

1 INTRODUCTION

With the recent rapid advances in computer and

network technologies, educational researchers have

developed methods, tools and environments for

computer-assisted learning (Hwang, 2002). Several

researchers have already addressed the importance

of adaptive learning, either in traditional forms of

instruction or in computer-assisted instruction. In

addition, several personalization techniques have

been proposed for developing web-based learning

systems (Santally and Alain, 2006).

In studying the effect of adaptive learning in

science courses, most researchers often pay attention

to the impact of a single type of personalization,

such as learning performance (including learner’s

profile and learning portfolio), learning style,

cognitive style of individual students, on the

determination of difficulty levels, learning paths or

presentation styles of subject materials (Triantafillou

et al. 2004). However, the interactions among

multiple sources of personalization information are

rarely taken into consideration.

An adaptive e-Learning system gives the learner

the opportunity to select learning materials or

contents according his/her style, profile, interest,

previous knowledge level. A number of works have

been conducted in the area of adaptive learning

(Kamceva and Mitrevski, 2012).

The study of how learners learn has been a

concern for researchers for many years [Pinto, et al.

2008]. In traditional classroom system, an instructor

can control this aspect based on what s/he sees of

her/his learners’ reaction. However, for e-learning to

be effective, it should be adapted to one’s personal

learning style (Villaverde et al. 2006). Traditional e-

learning systems provide the same materials to all

learners. E-learning systems should be capable of

adapting the content of courses to the individual

characteristics of learners. Adaptive e-learning

systems attempt to address this challenge by

changing the presentation of material to suit each

individual learner. They collect information about

learner’s goals, preferences and knowledge in order

to adapt the education needs of that learner.

An e-learning system must be based on learner’s

learning style which makes e-learning more

effective and efficient. However, most e-learning

systems do not consider learner characteristics. One

of the most desired characteristics of an e-learning

system is personalization, as people with different

skill sets use the system. This paper presents an

adaptive e-learning system based on the learner’s

learning style and preferences. The system identifies

the learner’s learning styles tendency through an

initial assessment test. The test’s score will be used

by the system as a basis to provide the learner a

presentation of learning materials more closely to

Chrysoulas, C. and Fasli, M.

Building an Adaptive E-Learning System.

DOI: 10.5220/0006326103750382

In Proceedings of the 9th International Conference on Computer Supported Education (CSEDU 2017) - Volume 2, pages 375-382

ISBN: 978-989-758-240-0

375

his/her knowledge level. This is the first input of the

user in the system in parallel with some basic

information that they provide in the form of a

profile. In this paper, a multi-source adaptive

learning system is also proposed. The proposed

system can easily construct adaptive subject series of

tests and propose sources for the learners to study

and tutors to talk to, by taking both student learning

behaviours and learning styles as part of the

personalization information.

The rest of the paper is structured as follows:

Section 2 gives a brief overview of the abstract

Adaptive e-learning architecture and an in depth

insight on how the User profile is modelled in the

proposed system. Section 3 discusses what a

learning path is and its importance. Section 4

discusses the adopted machine learning approach in

order to put the needed personalization and

adaptivity into the system. Section 5 presents some

initial results of the analysis while Section 6

concludes the work.

Figure 1: Adaptive E-learning System Architecture.

2 USER PROFILE DESIGN

As depicted in Figure 1, the system is using the

intelligence coming from the machine learning part,

to dynamically propose learning paths to individual

learners, based on their requirements and needs. The

learning paths are formed based on previous users,

with similar characteristics, interactions with the

system. Learning paths that led to successful

outcome, meaning the leaner improved his/her skills

are targeted. The Machine Learning part is

responsible for accessing the dataset that is extracted

from the RDF database and for feeding it to the

actual machine learning algorithm (see Section 4 for

more information). The usage of Resource

Description Framework (RDF) database (W3C RDF

Working Group, 2014) makes it easy for the system

to update the user’s profile thus keeping track of the

whole learning process. The Recommendation part

is also a sub-component of the Machine Learning

part. The personalized recommendation(s) provided

are the outcome of the machine learning interference

with the system. Association analysis based on the

Apriori (Agrawal and Srikant, 1994) algorithm is

taking place.

We use the terms “user” and “learner”

interchangeably to refer to the same entity. An

“active” user/learner is the one that we are currently

considering or dealing with. In order to be able to

guide the user through the learning and assessment

process, information about the user and his/her

activities will need to be collated and recorded in a

user profile. The User Profile will need to record

both static and more dynamic information. The User

Profile as a cornerstone component of the proposed

E-learning system is well studied and documented

during the development process. Figure 2 provides a

schematic overview of the proposed User Profile.

Figure 2: User Profile Modelling.

2.1 Security and Role

Security information is about the user’s authority to

use the system and it comes in the form of: (i)

Username; and (ii) Password. OAuth 2.0 (OAuth

2.0, 2012) protocol is providing the needed security

for the users to securely sign-in to the system.

Role information provides an insight in the

relationship among the system’s users, and can be

described as: (i) Administrator; (ii) Tutor; and (iii)

Student/Learner. In this way, the system is in

position to easily provide different services to

different type of users.

CSEDU 2017 - 9th International Conference on Computer Supported Education

376

2.2 Personal Information

This is predominantly static data recording some

basic information on the user: (i) Name; (ii)

Address; (iii) Phone; (iv) School; (v) Email; (vi)

Sex; and (vii) Telephone. There may be useful

pieces of information such as the postcode that could

help identify other users living in the locality that the

active user may wish to connect to and the school

that the user/learner is attending as this may be

useful in identifying/putting him/her in a group or

further on looking at class or even school level

performance metrics and the performance of the

active user within a class/school.

2.3 User Interests

User interests in the system essentially represent the

topics that the learner is working on (or wishes to

work on) and improve his/her performance. There

are two ways that user interests can be collected:

implicitly and explicitly. Implicitly capturing user

interests would entail that the user behaviour (topics

chosen to read, or specific tests chosen that cover

specific topics) would need to be observed

(unobtrusively) and then these mapped against the

system’s database (RDF database) by using semantic

similarity measures (Slimani, 2013). The explicit

way of recording user interests would entail that at

the time of registration and then periodically, the

user would explicitly indicate his/her interests in

topics drawn from the database. In other words, the

user needs to be shown parts of the database

capturing the topics and choose from these.

User interests are declared in advance by the user

and hence captured explicitly. In this way, the user

interests then simultaneously indicate the learning

objectives and therefore in a way what needs to be

achieved by the user. So for instance, an interest in

fractions means that the user wants to master the

topic of fractions.

2.4 Performance/Test Data

The dynamic data within the user profile are in

essence the data generated from the user taking tests.

Information like: (i) Test id; (ii) Overall score; (iii)

Date taken; (iv) Time to completion; (v) Qx-id,

score (or simply correct/incorrect).

We may wish to record the time it takes a user to

complete a question as a) this may be different from

user to user; b) it can be used to distinguish between

difficult and more easy questions (and even use this

information later on to adjust the level of difficulty

of a question). Also the level of difficulty of the

individual questions involved in the test can give an

insight on the overall level of difficulty of the test.

The data generated from the tests will be used to

capture and record the user’s progress on a topic. A

test can have multiple topics covered through the

questions.

2.5 Log Data and Actions within the

System

As the user interacts with the e-learning system, s/he

is doing so by performing a set of actions. As the

user logs in with a unique ID, therefore his/her

activity can be tracked. We would presume that

these log data are “raw”.

Assuming that activity will be recorded in

sessions, the raw data would look like: (i) Session

ID/UserID; (ii) Date/Timestamp; (iii) Duration; and

(iv) Action x, timestamp x. Where Action can be:

1. Test_Taken, TestID;

2. Topic_Browsed, TopicID

3. Topic_searched, TopicID

4. Talked to a Tutor, TutorID

We can make a distinction in the educational

platform between self-directed and directed learning

(Brookfield, 2009). The actions we may wish to

record vary somewhat between these two, although

they have many elements in common. In both areas,

the concept of “engagement” is very important.

Engagement could be measured by a combination of

the following:

(A) Self-directed: (i) How often one logs into

the system; (ii) Session duration; (iii) Page view

duration; (iv) Abandoned tests; (v) Results review

(has the learner always reviewed results?); (vi)

Following links; and (vii) Repeating topics, that is,

taking another test in the same topic.

(B) Directed: (i) How often the learner has

contacted the educator; (ii) The feedback the

educator has given; (iii) The additional tests the

educator has assigned; (iv) Whether the user has in

fact taken these or not; and (v) Links that the

educator has recommended.

Such “actions” (or in other words, how is it that

the user interacts with the system) would be

important as they would tell us what users do and

helps us identify learning paths (see Section 3) by

aggregating either a specific user’s actions or the

actions of multiple users. In other words, the log

data could be mined to identify actions of individual

users and/or groups and distinguish between

successful learning paths (or sequences of actions)

and not so successful learning paths.

Building an Adaptive E-Learning System

377

3 LEARNING PATH

The end objective is for learners to master specific

topics and the ways they can assess their own

progress is by taking tests, talking to tutors, and

visiting educational sources (webpages) proposed by

the system or/and tutors. There are two aspects of

the learning path: the modelling of what a learning

path is and then its extraction. The modelling of

what a learning path is and its conceptual meaning is

important. Then what follows is the actual extraction

of the learning path in terms of SPARQL queries

(W3C SPARQL Query Language for RDF, 2013).

A learning path is not just a series of tests taken,

but it is a series of actions that lead to a successful

outcome. A learning path expresses a set of actions

as taken by a learner in relation to a specific topic

that needs to be mastered (or in order to achieve a

threshold of performance in a specific topic). A

learning path can be extracted with SPARQL queries

from the RDF database. A learning path is all of the

actions that are related to a topic or a combination of

topics, culminating in the successful achievement of

the outcome, measured by a test.

All of these data points and their attributes are

used by the machine learning to identify the

attributes of successful learning paths. Whether or

not the learning path is successful is indicated by the

answer to. The following subsection 3.1 presents in

depth two use cases of how a learning path is

constructed and who the involved parties and roles

are.

3.1 Use Cases

Initially a user (Learner 1, from now on), logs to the

system. Subsequently an initial Test is proposed to

Learner 1 in order to estimate the level of his/her

knowledge. Based on that, the system may propose

some Links for extra studying, some tutors to speak

to, or even some extra tests. After this initial phase,

let’s assume that Learner 1 chooses Topic 1 and

takes a Test T1. He scores 30. From the score

achieved (and assuming that 40 or 50 is a threshold

of performance that needs to be achieved), it is

obvious that Learner 1 requires extra support and

help in Topic 1.

What should follow then is an initial

recommendation on improving knowledge on Topic

1. The recommendation can be based i) on what

actions similar learners followed as actions and

appear to have helped them in succeeding in their

tests; ii) on what the specific topic that a learner did

poorly in is about (for example Topic 1 depends on

understanding and mastering two subtopics 1.1 and

1.2 which are essential for answering questions on

Topic 1); and iii) a combination of (i) and (ii).

So Learner 1 visits page P1 and page P2

associated with Topic 1, and speaks to a tutor or

both. After that, Learner 1 takes Test T2 in Topic 1.

Learner 1 achieves a score of 60. Learner 1 managed

to improve the score in Topic 1 (which was the

target). Clearly something that has happened

between taking tests T1 and T2 has led Learner 1 to

improve his/her performance. This could be the

result of the recommendation made by the system

(see (i)-(iii) above) or Learner 1 engaging with other

Learners or following suggestions for additional

study and tests by the tutor.

Therefore it is evident that it would be helpful to

them, if the system is able to provide some useful

hints of what previous Learners have done (tests

taken, tutors they spoke, etc.) and how they

performed when they interacted with the system.

The Machine Learning part will provide the system

with a pool of possible questions, tutors and links to

be followed by a Learner based on the interactions

that previous Learners of the system have had. These

suggestions will be per Topic.

Two Use-cases/Scenarios have been spotted

where the system can propose – interact with the

Learner. The first one is just after the Learner has

logged-in into the system, in order to take an initial

assigned test (assuming that there is a need for that),

and the second one after finishing the initial test, and

the Learner proceeds to the choice of the Topic to

study.

Use-case: Initial Assi

ned Test

ID:1

Brief description:

This use-case describes how the initial

assigned test is provided to the Learner, and the

options following the completion of it.

Primar

actors:

Learner

Secondar

actors:

Platform, Tutors

Preconditions:

Have securely logged-in into the System

Main flow:

a) Learner chooses Topic x

b) The System proposes an initial test to

check the level of the Learner

CSEDU 2017 - 9th International Conference on Computer Supported Education

378

c) The System checks the level of the

Learner

d) The System provides the learner with

information like tutor(s) to talk to, webpages to

visit, etc., based on similar behaviour of previous

Users/Learners. Can also provide non-system

actions like, practice more, spend more time on

studying, etc.

e) The Learner may follow or not the

suggestions provided by the system

Post conditions:

None

Alternative flows:

It is possible to take the assigned test before

choosing the Topic.

Use-case: Learner chooses a Topic to

ID:2

Brief description:

This use-case describes the steps a Learner is

following from choosing a topic to study till the

completion of his learning session/process.

Primar

actors:

Learner

Secondar

actors:

Platform, Tutors

Preconditions:

The Learner must complete the initial

assigned test first (if it is considered a hard

condition)

Main flow:

a) Learner chooses Topic x

b) The system provides them with

information like tutor(s) to talk to,

webpages to visit, etc., based on actions

of past Users/Learners that have been

identified as having gone through

successful learning paths

c) Learner may follow or not the

suggestions

d) Learner chooses a Test to take. The

suggestions produced out of the previous

step (b) would probably be desirable to

remain visible on the screen (maybe on

the right side)

e) Learner finishes the Test

f) The System based on how the Learned

did in the Test, can propose new Tests,

Tutors to speak to and websites for the

Learner to follow

g) Learner chooses another Topic. New

suggestions based on actions of past

Users/Learners that have been identified

as having gone through successful

learning paths previously (and similar

behaviour) are suggested to the Learner

h) Learner ends the learning activity

Post conditions:

None

Alternative flows:

None.

An activity diagram presenting and visualizing

the aforementioned two Use Cases can be found in

Figure 3.

Figure 3: Use-Cases Activity Diagram.

4 MACHINE LEARNING

APPROACH

The Query Results Management (QRM)

component/module is responsible for managing the

data that are extracted from the queries to the RDF

database and for assembling the dataset that will be

fed to the algorithm. In Figure 4, an illustration of

where the QRM manager component is situated in

relation to the whole educational system and to the

Machine Learning component is presented. The

QRM component, a fundamental component is

responsible for supporting the following

functionalities: (i) Establishing a safe connection to

the RDF database; (ii) Querying the RDF database,

receiving the data; and (iii) Saving the data in a file

and in the proper format for the Machine Learning

Management (MLM) component.

Building an Adaptive E-Learning System

379

Figure 4: Component based System Architecture.

The Machine Learning Management (MLM)

component/module is responsible for accessing the

dataset that is formed from the Query Results

Management component (QRM) and for feeding it

to the actual machine learning algorithm. In Figure

4, an illustration of where the MLM manager

component is situated in relation to the whole

educational system and to the Query Results

Management component is presented. The MLM

component will be responsible for supporting the

following functionalities: (i) Accessing the dataset

formed by the QRM component; (ii) Parsing the

dataset, thus filtering out the un-successful learning

paths; (iii) Feeding the filtered dataset to the

machine learning algorithm; and (iv) Saving the

results coming out from the algorithm, to be used for

further visualization purposes. This sequence of

actions is described in more detail in the following

subsections.

4.1 Query Results Management

Component

Figure 5 presents the interactions of the QRM

module with the RDF Database and the Machine

Learning Module.

Figure 5: QRM Component Interactions.

QRM will communicate with the RDF Database

to query it and get the results. The results could be in

XML, JSON, CSV or TSV format. The QRM

Manager Component will be responsible to Securely

Accessing RDF Database by setting up a two way

Secure Sockets Layer (SSL) connection to the

Apache Jena Fuseki server [Fuseki, 2016] in order to

securely query the RDF database.

4.2 Machine Learning Management

Component

The QRM component is responsible for the two first

steps of the aforementioned procedure. The MLM

component will be responsible for the three last

steps of the described procedure. Figure 6 presents

the interactions of the MLM component with the

QRM Component (and the educational platform/

system).

Figure 6: MLM Component Interactions.

The MLM Component will support the following

functionality:

Accessing/Parsing the Dataset: After securely

connecting to the Apache Jena Fuseki server, and

executing SPARQL query(ies) against the endpoint,

we are obtaining a dataset containing the interactions

the learners had with the system. In the next step, the

MLM accesses the dataset and parses it in order to

filter out the unsuccessful learning paths. The

filtered dataset would only contain the learning paths

where the number of correct answered items

(questions) are more than the incorrect ones. For the

current implementation meaning that only the

over/equal to 50 scored learning paths will qualify.

Feeding the ML Algorithm: After filtering out the

unsuccessful learning paths, the dataset is ready to

be fed to the machine learning algorithm. The

machine learning algorithm used in the presented

work is the Apriori (Agrawal and Srikant, 1994),

since it deals well with datasets containing both

numerical and categorical values. The Apriori

algorithm is responsible for: (i) finding frequent

itemsets, and (ii) mining association rules from the

extracted itemsets. See Sections 4.3 and 4.4 for more

details.

Saving the results: In this step, the MLM

component saves the output of the machine learning

algorithm, so as to be used, for recommendation,

visualisation and/or other purposes.

The pseudocode of the Apriori algorithm is

presented in Algorithm 1.

CSEDU 2017 - 9th International Conference on Computer Supported Education

380

Algorithm1. The Apriori algorithm.

: Candidate itemset of size k

: frequent itemset of size k

(1) L

= {frequent items};

(2) for (k = 1; L

!= ∅; k++) do begin

(3) C

k+1

= candidates generated from L

;

(4) for each transaction t in database do

(5) increment the count of all candidates in

(6) C

k+1

that are contained in t

(7) L

k+1

= candidates in C

k+1

with min_support

(8) end

(9) return ∪

;

4.3 Finding Frequent Itemsets

The support of an itemset is defined as the

percentage of the dataset that contains this frequent

itemset. Frequent itemsets are a collection of items

that frequently occur together. In our specific case

an itemset is having the following format:

{userId, learningEventUri, leraningEventType,

difficultyLevel, contentUri, topicUri,

nextlearningEventUri, timestamp}

Support applies to an itemset, so we can define a

minimum support and get only the itemsets that

meet that minimum support. Support can range from

0 to 1. The confidence is defined for an association

rule like {Learner 1} ➞ {Topic 1}. The confidence

for this rule is defined as support ({Learner 1, Topic

1})/support ({Learner 1}). The support and

confidence are ways someone can quantify the

success of our association analysis. Let us assume

we want to find all sets of items with a support

greater than 0.6. We could generate a list of every

combination of items and then count how frequently

these occur.

4.4 Mining Association Rules from the

Extracted Itemsets

Association rules suggest that a strong relationship

exists between two items. From the dataset we have,

if we have a frequent itemset, {Learner 1, Question

1, Topic 1}, one example of an association rule is

Topic 1 ➞ Question 1. This means if someone

chooses Topic 1 Question 1, then there is a

statistically significant chance that the Learner will

answer Question 1. The converse does not always

hold.

In Section 4.3, an itemset is quantified as

frequent if it met our minimum support level. There

is a similar measurement for association rules. This

measurement is called confidence. The confidence

for a rule P ➞ H is defined as support (P | H)/

support (P). Similarly to the frequent itemsets

generation in Section 4.3, we can generate many

association rules for each frequent itemset. It would

be desirable if we could reduce the number of rules

to keep the problem tractable. We can observe that if

a rule does not meet the minimum confidence

requirement, then subsets of that rule also will not

meet the minimum. We can use this property of

association rules to reduce the number of rules we

need to test.

5 RESULTS

A series of tests performed in order to check the

validity of the proposed approach. Python used for

the implementation of the machine learning

proposed approach/architecture. The outcome of the

association and data analysis that took place has

shown that the proposed framework is in position to

provide an insight on the behaviour of the students,

meaning how they interacted with the system and

spot common patterns that lead or not to successful

completion of a learning path:

1. Users and Learning Paths: The results contain

the full information of the users that followed a

successful or unsuccessful learning path. The format

is: {userId, learning EventUri, leraningEventType,

difficultyLevel, contentUri, topicUri,

nextlearningEventUri, timestamp}.

2. Successful Users and type of interaction per

topic: The results contain the full information

describing how each successful user answered the

questions and interacted with the system (talked to

tutor or/and followed links, etc.) per topic. The

format is: {userId, leraningEventType, topicUri}.

3. Percentage for each question per topic: The

results contain the information describing how each

question x belonging to topic y was dealt by

successful or not users. The format is: {topicUri,

contentUri, Percentage}.

4. Percentage for each topic, and interactions

involved: The results contain the information

describing how successful or not users did it per

topic, how many tutors were called and how many

links were followed. {topicUri, percentage,

NumberOf Tutors, NumberOfFollowedLinks}.

6 CONCLUSIONS

The paper presents an adaptive e-learning system

based on well defining the explicit characteristics of

the individual user/learners. The system identifies

Building an Adaptive E-Learning System

381

the learner’s learning styles through an initial

assessment test and some initial information a user is

providing during sign-in. The test score is used by

the system to complement the learner with the

needed material for him to successfully complete his

learning path.

The paper investigates the detection of E-

learners’ preferences within learning style

dimensions and showing relationship between

identifying the personality and learning materials

presentation. It contributes how to develop e-

learning to different learning styles and combine the

advantages of learning management systems with

those of adaptive systems. The experiments were

performed with 300 learners to show the impact of

learning styles on learners’ preferences. Several

patterns were found where learners with different

starting points managed to evolve based on the

proposed machine learning based recommendation

system.

Future work has to be two-fold: perform

experiments with significantly larger datasets, and

try to enrich the proposed system with more

machine learning methods (e.g. neural nets (NN),

support vector machines (SVNs), etc.) thus

enhancing the recommendation efficiency of the

whole system.

ACKNOWLEDGEMENTS

The work was supported by the «DADO: Data

Analytics Driven by Ontologies» project, a

collaborative research and development project

funded by Innovate UK.

REFERENCES

Agrawal, R., Srikant, R. (1994). Fast algorithms for

mining association rules, in: Proc. 20th Int. Conf. Very

Large Data Bases, VLDB, 1994, pp. 487– 499.

Brookfield, S. (2009). Self-directed learning. In R.

Maclean & D. N. Wilson (Eds.), International

handbook of education for the changing world of work

(pp. 2615–2627). Dordrecht: Springer.

http://dx.doi.org/10.1007/978-1-4020-5281-1.

Fuseki: serving RDF data over HTTP. (2016)

https://jena.apache.org/documentation/serving_data/

Hwang, G. J. (2002). On the development of a cooperative

tutoring environment on computer networks. IEEE

Transactions on System, Man and Cybernetic Part C,

32(3), 272–278.

Kamceva, E., Mitrevski, P. (2012). On the General

Paradigms for Implementing Adaptive e- Learning

Systems”, ICT Innovation Web Proceedings ISSN

1857-7288, pp.281-289.

OAuth 2.0: The OAuth 2.0 Authorization Framework

(2012). https://tools.ietf.org/html/rfc6749.

Pinto, J., Ng, P., Williams, S. K. (2008). The Effects of

Learning Styles on Course Performance: A Quantile

Regression Analysis. Franke College of Business,

Working Paper Series – 08-02.

Santally, M. I., Alain, S. (2006). Personalisation in Web-

based learning environments. International Journal of

Distance Education Technologies, 4(4), 15–35.

Slimani, T. (2013). Description and Evaluation of

Semantic Similarity Measures Approa-

ches. International Journal of Computer

Applications 80(10):25-33.

Triantafillou, E., Pomportsis, A., Demetriadis, S.,

Georgiadou, E. (2004). The value of adaptivity based

on cognitive style: an empirical study. British Journal

of Educational Technology, 35(1), 95–106.

Tsai, C.-C. (2004). Beyond cognitive and metacognitive

tools: The use of the Internet as an ‘‘epistemological’’

tool for instruction. British Journal of Educational

Technology, 35, 525–536.

Villaverde, J.E., Godoy, D., Amandi, A. (2006).

“Learning styles” recognition in e-learning

environments with feed-forward neural networks.

Journal compilation and Blackwell Publishing Ltd, pp.

197-206.

W3C Resource Description Framework Working Group

(2014). https://www.w3.org/ standards/techs

/rdf#w3c_all.

W3C SPARQL Query Language for RDF (2013).

https://www.w3.org/TR/rdf-sparql-query/

CSEDU 2017 - 9th International Conference on Computer Supported Education

382