Recommendation of Educational Content to Improve Student

Performance: An Approach based on Learning Styles

Alana Oliveira

1,3 a

, Mario Meireles Teixeira

2 b

and Carlos S. Soares Neto

2 c

Computer Engineering, Federal University of Maranh

ao (UFMA), Brazil

Department of Informatics, Federal University of Maranh

ao (UFMA), Brazil

PhD Program in Computer Science, DCCMAPI / UFMA, Brazil

Keywords:

Learning Analytics, Educational Data Mining, Content Recommendation, Learning Styles.

Abstract:

Virtual learning environments are a powerful tool in the teaching-learning process and can provide a variety of

utilization data that can be explored by data mining techniques to improve the understanding of student behav-

ior and performance. By using Learning Analytics, it is possible to identify potential problems, such as student

dropout or failures before they become irreversible, and indicate corrective actions to be taken by teachers. In

this context, content recommendation plays a prominent role since choosing the proper content for a certain

audience may motivate them to become more involved in the learning process. However, in distance educa-

tion settings nowadays, teachers do not know their students, thus it becomes difﬁcult to select the content most

suitable to their needs. In this paper, we propose a content recommendation architecture that takes into account

the learning proﬁle of students enrolled in an LMS to customize content recommendations to each learner’s

style. A proﬁle assessment tool, based on the Honey-Mumford learning style taxonomy was implemented and

some preliminary data obtained. We devised a recommendation scheme that considers the euclidean distance

between students’ learning styles when suggesting content to be studied. Our preliminary results indicate this

approach may be beneﬁcial to improve the teaching-learning process and student performance as a whole.

1 INTRODUCTION

Learning Management Systems (LMSs) have been

considered a valuable tool in the teaching-learning

process, being used mainly in distance learning but

also in formal teaching. Such tools are intended to

provide content to students to assist them in studying

a particular subject. Coupled with this, these environ-

ments have the potential to generate a wealth of data

about their utilization by students and teachers, which

can be explored to help understand students’ behavior

and performance aspects.

Educational Data Mining is a ﬁeld of research that

has as its object of study such aspects, using statis-

tical and machine learning methods, and techniques,

to explore and analyze data originating in an educa-

tional context in order to better understand the dif-

ferent variables involved in this complex process of

teaching-learning mediated by virtual tools (Romero

https://orcid.org/0000-0001-7870-3943

https://orcid.org/0000-0001-8771-1478

https://orcid.org/0000-0002-6800-1881

and Ventura, 2010).

An emerging area in this context is Learning An-

alytics, which deals with the study of technology-

mediated learning and has its roots in diverse ar-

eas such as educational data mining, business intel-

ligence, web data analysis, and recommendation sys-

tems. More speciﬁcally, Learning Analytics refers to

the measurement, collection, analysis and presenta-

tion of data about learners and their contexts in order

to understand and optimize the learning process and

the environments in which it takes place (Baker and

Inventado, 2014) (Ferguson, 2012).

In face-to-face teaching, adjustments in the

teaching-learning process occur based on teachers’

experience in analyzing student feedback and taking

corrective actions that they deem appropriate. How-

ever, teachers’ performance has been increasingly

hampered by the increasing size of the classes, given

the recent massiﬁcation of education.

In the current context of education mediated by

learning management systems, this difﬁculty is even

more noticeable, since the amount of students en-

rolled in classes becomes overwhelming and eventu-

Oliveira, A., Teixeira, M. and Neto, C.

Recommendation of Educational Content to Improve Student Performance: An Approach based on Learning Styles.

DOI: 10.5220/0009436303590365

In Proceedings of the 12th Inter national Conference on Computer Supported Education (CSEDU 2020) - Volume 2, pages 359-365

ISBN: 978-989-758-417-6

359

ally overloads the instructor. Thus, he cannot pro-

vide personalized attention to each student and may

lose track of each individual’s context. Thus, new

approaches are welcome to identify problematic sit-

uations in the teaching-learning process at their in-

ception, such as student potential to dropout or fail a

course.

One key aspect of this process is content recom-

mendation (Mendes et al., 2017). It aims to pro-

vide educational material to students to increment

their knowledge about a subject and their chances of

succeeding. However, simply suggesting content to

students as if they were a homogeneous body may

not yield the intended results since each student has

their own learning style. In this paper, we develop a

web-based tool to collect student data via a question-

naire based on the Honey-Mumford taxonomy, aim-

ing to identify students’ learning proﬁles. This in-

formation serves as input to a recommendation sys-

tem that assists teachers and students in choosing the

better-suited activities or materials for distinct learn-

ing styles in a class. We have performed some prelim-

inary validation that indicates promising results yet to

be measured.

The remainder of this paper is organized as fol-

lows: Section 2 deals with the main theoretical as-

pects related to this research. Section 3 proposes a

methodology, mediated by an LMS, to help teachers

follow their students’ progress within a course, iden-

tify potential problematic situations and take appro-

priate actions to remedy them. To this end, a recom-

mender architecture, based on student learning styles

is described in Section 4 and some preliminary results

presented in Section 5. Section 6 summarizes the pa-

per and indicates some directions for future work.

2 THEORETICAL BACKGROUND

2.1 Learning Styles

Each person is a unique being, differing from others

in the way they think, act and relate to others. Sim-

ilarly, each human being has their own way of learn-

ing, those who learn by doing and those who observe,

those who are multitasking, and those who need to

focus on one task at a time.

Several studies intend to classify students accord-

ing to the way most favorable to their learning, that is,

according to their learning style, seeking to identify

the best way for a student to assimilate the knowledge

that is transmitted.

For (Kolb, 2014), learning is the process by which

knowledge is created through the transformation of

experience. And knowledge is not something that can

simply be transmitted or acquired, it is the result of

a process and can be created and recreated continu-

ously. Kolb also believes that people can be classiﬁed

according to their way of learning into learning styles

(or preferences) as diverging, converging, assimilat-

ing, and accommodating. This classiﬁcation can be

used to provide teachers with information so that they

consider the best way their students can learn and thus

be able to achieve better success in their teaching.

Honey and Mumford have identiﬁed, based on

Kolb’s work, four learning styles or preferences: Ac-

tivist, Theorist, Pragmatist, and Reﬂector. The au-

thors recommend that, to maximize personal learning,

each student should know their own learning style and

then look for opportunities to learn using that style.

In order to assist the student in obtaining their

learning style, Honey and Mumford have developed

a Learning Style Questionnaire (Honey, 2001). This

questionnaire consists of 80 yes/no answer questions,

where the respondent points out as true the statements

to which he agrees. At the end, a score is obtained for

each style (Activist, Theorist, Pragmatist and Reﬂec-

tor). This score refers to the student’s intensity in that

proﬁle, and may range from Very low preference to

Very strong preference for each of them.

Activists are people who learn by doing. They

learn best when engaging in new experiences and

working with others to solve problems, games and

simulations of real situations. Reﬂectors learn by

watching and thinking about what is happening. They

like to consider all the implications before giving an

opinion. Theorists like to understand the theory be-

hind actions. They like to analyze and synthesize, and

they are uncomfortable with subjectivity. And ﬁnally,

Pragmatists like to try things out. They like new ideas

that can be put into practice.

The VARK model is another approach to learning

style classiﬁcation proposed by (Fleming and Baume,

2006). This model is based on classifying learners

into four modalities, or learning preferences, namely:

visual, aural, read/write and kinesthetic.

Such classiﬁcation is accomplished by applying

a form containing 64 questions, where answers will

be marked according to a student’s individual prefer-

ences. In the end, the form presents the summary of

the answers within each of the modalities, and the in-

dividual can be part of more than one modality.

Each modality has its own characteristics. For ex-

ample, aural learning is the learning style in which

one learns through the aural sense. An aural learner

has listening and speaking as the primary means of

learning. Behaviors such as repeating content aloud

for memorization, or even the preference for videos,

CSEDU 2020 - 12th International Conference on Computer Supported Education

360

podcasts, and participation in study groups are more

common in people of this style.

(Fleming and Baume, 2006) believe that the use

of questionnaires to deﬁne learning styles can be use-

ful, but its real value lies in the self-knowledge it can

generate to each person when analyzing the score ob-

tained. This classiﬁcation can serve as input for teach-

ers so that they can choose the most appropriate activ-

ities for each modality.

In (Bartle, 1996), another classiﬁcation of student

proﬁles aimed at games and gamiﬁed environments,

the so-called Bartle Archetypes, is proposed, which

may be helpful in composing student proﬁles and in

understanding the reasons behind high dropout rates,

for example. Bartle believes players are motivated by

the autonomy, challenges, relationships, and sense of

power provided by games, and knowing how to lever-

age these elements to maintain student motivation and

interest is extremely valuable.

Bartle proposes categorizing players into four pro-

ﬁles according to how they relate to and act in games.

Achievers are driven by the goal of the game and

strive to achieve it, accumulating wealth, scoring

points and collecting as many items as possible. Ex-

plorers are interested in ﬁnding out as much as pos-

sible about the game, ﬁnding secret passages and un-

raveling the entire game world; students with this pro-

ﬁle prefer to know all the paths or stages they should

go through before beginning the journey. Socializers

prefer to relate to other players, even outside the game

environment; and ﬁnally, Killers are concerned with

asserting their existence in competition with other

players or the environment.

If teachers can better identify and understand their

students learning styles, they will be able to provide

better educational experiences to them and lead each

individual on a more efﬁcient path according to their

proﬁle.

In this paper, we use the aforementioned Honey-

Mumford questionnaire to assess students’ proﬁles

in order to assist them in selecting educational con-

tent better suited to their learning preferences, as de-

scribed in Section 4.

2.2 Educational Data Mining

Educational Data Mining (EDM) uses traditional and

statistical machine learning methods and techniques

to explore and analyze data obtained from educational

contexts. The main objective of this approach is to an-

alyze the different variables involved in the teaching-

learning process and use them to develop predictive

models in order to classify students according to their

performance (Fernandes et al., 2019).

EDM makes use of educational systems databases

to understand students and their learning styles more

thoroughly in an effort to devise educational strate-

gies that will increase their academic achievement and

success rate at the end of each term.

According to (Chalaris et al., 2014) based on

(North, 2012), EDM consists of a process divided into

six steps or phases: business understanding, data un-

derstanding, data preparation, modeling, evaluation,

and deployment. The business understanding step

consists in understanding the objectives and require-

ments of the project. Understanding the data allows

us to identify any data quality issues as well as inter-

esting subsets for formulating non-obvious hypothe-

ses. Data preparation comprises necessary tasks such

as data cleaning, transformation, and selection, in or-

der to build the ﬁnal data set to be used in the model-

ing phase, where different techniques can be applied.

Modeling is followed by the Evaluation step, which

determines the feasibility of applying the model and

whether business objectives have been met. The ﬁnal

step is Deployment, which speciﬁes how to employ

the developed models and the actions to be taken.

There are many data mining techniques that can

be applied to educational data and each can provide

useful results that help solve many educational prob-

lems. Data mining tasks, such as grouping, can re-

veal overall student characteristics, while prediction

(classiﬁcation and regression) and relationship min-

ing (association, correlation, sequential mining) can

help teachers mitigate student dropout rate, retention,

increase individual performance and improve learning

outcomes. This can help provide a more customized

learning process, maximize the efﬁciency of the edu-

cation system, and reduce the cost of educational pro-

cesses (Zhang et al., 2010).

2.3 Recommender Systems

Recommender systems have been one of the major di-

rect applications of artiﬁcial intelligence in people’s

lives. Due to a large amount of data that is gener-

ated daily on the internet, it is common for systems

integrated into websites to ﬁlter products and services

with a higher chance of pleasing the user. They work

as a friend, who, by knowing your taste, recommends

websites, videos, movies, music, and products.

In such systems, recommendations are solely

data-driven, without human intervention, and users’

historical patterns are identiﬁed and analyzed to rec-

ommend products and services that the users them-

selves did not know they needed.

However, recommendations can be made not only

for selling products, but also for content such as

Recommendation of Educational Content to Improve Student Performance: An Approach based on Learning Styles

361

news, articles, multimedia content, and even tourist

attractions (Smirnov et al., 2016). Recommenda-

tions can be made based on what people with simi-

lar proﬁles might like, the so-called collaborative ﬁl-

tering (Sarwar et al., 2001), or based on user his-

tory, namely, content-based ﬁltering (Van Meteren

and Van Someren, 2000). Content-based ﬁltering is

based on ﬁnding content similar to what the user has

been consuming and then recommending new content

to them, while collaborative ﬁltering recommends

content that has been consumed by other users with

a similar proﬁle. In an educational context, recom-

mendation systems can be applied to e-learning tasks

for recommending resources such as articles, books,

podcasts, or videos to students with the aim to tutor

them and make them achieve better performance.

3 METHODOLOGY

In this research, we propose a methodology that al-

lows teachers to monitor the progress of students in a

course, identifying potential problem situations early

on, so that it is possible to offer differentiated treat-

ment to students who need it. Thus, we expect to

minimize problems such as demotivation, failure, and

student dropout, increasing the success rate of the

teaching-learning process as a whole.

We assume the course will be mediated by a

Learning Management System (LMS), in face-to-face

or distance education settings. LMSs often generate a

wealth of data about their use by students and teach-

ers that, if well explored, can become a valuable tool

for tracking student achievement.

In Figure 1, we present a general scheme of how

we believe the proposed methodology can intercon-

nect the different actors and entities of the teaching-

learning process. The LMS is the intermediary ele-

ment and teachers are responsible for providing con-

tent, tasks, quizzes, questionnaires and various ac-

tivities, which will be read, watched and solved by

students during a course. In this process of inter-

action of students and teachers with the LMS, us-

age statistics are generated, such as number of con-

tent views, comments made by students, percentage

of successes/failures for a task, number of attempts,

compliance with deadlines, among many other as-

pects. Those statistics are computed globally, by class

and individualized by student.

This Data Collection phase will generate educa-

tional datasets obtained from LMS use, which will

serve as input to an Educational Data Mining phase

where machine learning models will be used to pre-

maturely identify students who are likely to fail a

course, are candidates to drop out, disoriented as to

what content to study or which tasks to solve, among

other problems. Data mining will also enable stu-

dents to be classiﬁed into learning proﬁles so that they

can receive customized service. Finally, based on the

previous classiﬁcation, we come to an Intervention

phase, in which semi-automatic preventive and cor-

rective actions will be triggered by the system, and

validated by the teacher, in order to rescue a student

with problems and increase their likelihood of com-

pleting the course successfully.

One point to consider is that students are individu-

als with different characteristics, interests, and goals,

so a given content or assignment that seems attractive

or simple to one person can be boring or complicated

to someone else. Therefore, it is important to iden-

tify the Learning Proﬁle of the students enrolled in a

course, for example, by means of a questionnaire ap-

plied to each student when registering in the LMS.

Thus, content that is likely to be more attractive

to the proﬁle into which a given student ﬁts will be

recommended. Besides, if that student has any issues

during the course, he may receive corrective actions

that are more appropriate to his proﬁle. We believe

that adjusting the content and tailoring corrective ac-

tions to each student’s learning proﬁle or style are

essential, as it allows, in a context of such diverse

and numerous classes, to provide more personal, cus-

tomized treatment, as if the students had the teacher

by their side all the time. We even intend to adapt the

system interface to a student’s proﬁle, thus seeking to

provide a differentiated and potentially more motivat-

ing user experience.

This approach does not intend to dismiss the ﬁg-

ure of the teacher as a mediator and facilitator of the

teaching-learning process, quite the opposite. The

master will have even more tools at hand, stemming

from the methodology and products to be developed,

which will allow him to conduct the course more ef-

fectively, acting globally and also individually.

Another important point is that students’ proﬁle,

initially collected from the application of question-

naires, can and should be adjusted throughout their

history in a given course or LMS, thus inferring new

behaviors and preferences that the students them-

selves not even perceived as their own characteristics

until then. We consider this adequacy of the content

and corrective actions to the student proﬁle as an im-

portant contribution to be achieved in this research.

Although current technology allows the massiﬁcation

of educational content, reaching an increasing num-

ber of students, it fails when considering that every-

one learns and feels content the same way, what can

often contribute to failure.

CSEDU 2020 - 12th International Conference on Computer Supported Education

362

Figure 1: Overview of actors and entities in a teaching-learning process interacting by means of an LMS.

4 RECOMMENDER

ARCHITECTURE BASED ON

LEARNING STYLES

This paper is focused on recommending educational

content to students in order to help them achieve

greater success in their courses. This applies to stu-

dents who have had a good performance and may be

interested in augmenting their knowledge, and also

to individuals who have been experiencing difﬁcul-

ties and may need some extra material for self-study

and improvement.

Our Recommender module (Figure 1) uses as in-

put the datasets collected by the LMS in order to pro-

vide content recommendation using collaborative ﬁl-

tering. A well-known problem in recommender sys-

tems is the cold-start problem, which is related to low

precision when recommending content to a new user,

since the system has hardly any information about

their preferences. In order to circumvent this prob-

lem, the proposed recommender architecture lever-

ages previously obtained students’ learning style in-

formation in order to provide customized content rec-

ommendations, tailored to a student’s proﬁle. Thus, it

is possible to suggest more accurate and useful con-

tent to learners.

Let us assume a given student is having problems

with a UML modeling topic in a Software Engineer-

ing course. Since this student is new to the LMS, the

system has limited background information on their

preferences, considering they have not yet contributed

much by rating available content. In this case, the

system would end up recommending a generic, seem-

ingly adequate content to the student, that might or

might not be useful. In the latter case, the student

might feel demotivated and even discredit future rec-

ommendations made by the LMS.

But now suppose the LMS has additional infor-

mation about the student, concerning how much they

belong to a certain learning style according to the

Honey-Mumford taxonomy, for example. In the ques-

tionnaire we applied to the students there are four

groups of 20 questions each, aimed at assessing a per-

son’s conformity to one of the four styles, as speciﬁed

by the aforementioned researchers (Figure 2).

Figure 2: Result of an individual’s learning style according

to the Honey-Mumford approach.

In the case of a student who recently signed up to the

LMS, there is limited information about what content

might be useful to them. Consequently, the system

will try to predict their ratings to content they have

Recommendation of Educational Content to Improve Student Performance: An Approach based on Learning Styles

363

not yet used by calculating the similarity between the

new student and other students who have been in the

LMS for a longer period. The similarity in this case

will be computed as follows:

sim

activist − u

activist)

re f lector − u

re f lector)

theorist − u

theorist)

pragmatist − u

pragmatist)

(1)

Note that the similarity between users u

and u

given by the euclidean distance between their pref-

erences in each of the four learning styles: Activist,

Reﬂector, Theorist, and Pragmatist. In this way, we

can predict the evaluation P

u,i

a given user u would

give to content i using the formula:

u,i

∑

v,i

· sim

u,v

)

∑

sim

u,v

(2)

Here, r

v,i

is the rating provided by user v to content

i. This implies that content ratings provided by users

with similar learning styles will have stronger impact

on ratings predicted for new users of the same kind.

In Figure 3, we depict a general scheme of how our

approach should work.

Figure 3: Educational content recommendation weighted

by learning style similarity.

As the student progresses through the LMS, by inter-

acting and rating content, content recommendations

become more precise and tailored to their preferences

and learning style.

5 PRELIMINARY RESULTS

The proposed recommender architecture based on

learning styles still deserves further validation. It

has not been widely used in courses with a larger

group of students. Up to this point, our results have

been very promising and style-based recommenda-

tions seem rather adequate.

Figure 4 shows the dashboard for a user highlight-

ing his related peers and also some content recom-

mendations he might enjoy. Figure 5 shows the sys-

tem’s content catalog tagged by subject and learning

style.

Figure 4: User dashboard.

6 CONCLUSION

Educational Data Mining has been used as a tool for

decades in order to analyze data originated in edu-

cational environments and improve educators’ under-

standing of the different variables involved in such a

complex scenario. Learning Analytics, an emerging

discipline, collects and analyzes data about students

seeking to enhance the learning process and the envi-

ronments where it occurs.

This paper describes a high-level methodology for

students’ performance follow-up and ﬁne-tuning in

Learning Management Systems (LMSs) where data

about LMS usage is measured, collected, analyzed,

and used to make predictions about learners’ perfor-

mance and point out potential failures in the learning

process before they occur, suggesting corrective ac-

tions to be taken by the teacher and students.

An essential role is such a scenario is content rec-

ommendation. We advocate that the use of learning

CSEDU 2020 - 12th International Conference on Computer Supported Education

364

Figure 5: Content catalog.

proﬁles as additional information for picking out con-

tent can provide better content selections that will ful-

ﬁll students’ needs and expectations. In this paper,

we detailed a Recommender Architecture based on

Learning Proﬁles to be used in an educational set-

ting mediated by a Learning Management System.

Our recommender uses collaborative ﬁltering to se-

lect content similar to what a user has already studied

and leverages student proﬁle information to ﬁlter the

content most suitable to a user’s needs. Thus, peer

rating as well as proﬁle indications will guide content

selection and improve the learning process. In future

work, we even intend to adapt content by taking into

account the learner’s proﬁle.

Further validation should be performed in class-

room in the near future but our preliminary results

ﬁrmly indicate we are in the right track.

REFERENCES

Baker, R. S. and Inventado, P. S. (2014). Educational data

mining and learning analytics. In Learning analytics,

pages 61–75. Springer.

Bartle, R. (1996). Hearts, clubs, diamonds, spades: Players

who suit muds. Journal of MUD research, 1(1):19.

Chalaris, M., Gritzalis, S., Maragoudakis, M.,

Sgouropoulou, C., and Tsolakidis, A. (2014).

Improving quality of educational processes provid-

ing new knowledge using data mining techniques.

Procedia-Social and Behavioral Sciences, 147:390–

397.

Ferguson, R. (2012). Learning analytics: drivers, develop-

ments and challenges. International Journal of Tech-

nology Enhanced Learning, 4(5/6):304–317.

Fleming, N. and Baume, D. (2006). Learning styles again:

Varking up the right tree! Educational developments,

7(4):4.

Honey, P. (2001). Honey and mumford learning styles

questionnaire. Peter Honey Learning. Retrieved April,

25:2007.

Kolb, D. A. (2014). Experiential learning: Experience as

the source of learning and development. FT press.

Mendes, M. M., Carvalho, V. C., Ara

ujo, R. D., Dorc¸a,

F. A., and Cattelan, R. G. (2017). Clustering learning

objects in the ieee-lom standard considering learning

styles to support customized recommendation systems

in educational environments. In 2017 Twelfth Latin

American Conference on Learning Technologies (LA-

CLO), pages 1–8. IEEE.

North, M. (2012). Data mining for the masses. Global Text

Project Athens.

Romero, C. and Ventura, S. (2010). Educational data min-

ing: a review of the state of the art. IEEE Transactions

on Systems, Man, and Cybernetics, Part C (Applica-

tions and Reviews), 40(6):601–618.

Sarwar, B. M., Karypis, G., Konstan, J. A., Riedl, J., et al.

(2001). Item-based collaborative ﬁltering recommen-

dation algorithms. Www, 1:285–295.

Smirnov, A. V., Ponomarev, A., and Kashevnik, A. M.

(2016). Tourist attraction recommendation service:

An approach, architecture and case study. In ICEIS

(2), pages 251–261.

Van Meteren, R. and Van Someren, M. (2000). Using

content-based ﬁltering for recommendation. In Pro-

ceedings of the Machine Learning in the New Infor-

mation Age: MLnet/ECML2000 Workshop, pages 47–

56.

Zhang, Y., Oussena, S., Clark, T., and Kim, H. (2010). Us-

ing data mining to improve student retention in higher

education: a case study. In In International Conerence

on Enterprise Information Systems. Citeseer.

Recommendation of Educational Content to Improve Student Performance: An Approach based on Learning Styles

365