Eighth IEEE International Conference on Advanced Learning Technologies

Identifying Learning Styles in Learning Management Systems by Using

Indications from Students Behaviour

Sabine Graf * Kinshuk Tzu-Chien Liu

National Central University Athabasca University National Central University
Graduate Institute of Learning School of Computing and Graduate Institute of Learning
and Instruction, Taiwan Information Systems, Canada and Instruction, Taiwan
[email protected] [email protected] [email protected]

Abstract factor in the learning process and agree that

incorporating them in education has potential to
Making students aware of their learning styles and facilitate learning for students.
presenting them with learning material that Learning styles can be considered in different ways.
incorporates their individual learning styles has A first step is to make learners aware of their learning
potential to make learning easier for students and styles and show them their individual strengths and
increase their learning progress. This paper proposes weaknesses. The knowledge about their learning styles
an automatic approach for identifying learning styles helps students to understand why learning is
with respect to the Felder-Silverman learning style sometimes difficult for them and is the basis for
model by inferring their learning styles from their developing their weaknesses.
behaviour during they are learning in an online Furthermore, students can be supported by
course. The approach was developed for learning matching the teaching style with their learning style.
management systems, which are commonly used in e- Providing students with learning material and activities
learning. In order to evaluate the proposed approach, that fit their preferred ways of learning can make
a study with 127 students was performed, comparing learning easier for them. This matching hypothesis is
the results of the automatic approach with those of a supported by educational theories. Moreover, studies
learning style questionnaire. The evaluation yielded such as those by Bajraktarevic, Hall, and Fullick [4]
good results and demonstrated that the proposed and Graf and Kinshuk [5] demonstrated supportive
approach is suitable for identifying learning styles. By results.
using the proposed approach, students learning styles For considering learning styles in education, the
can be identified automatically and be used for students learning styles need to be known first.
supporting students by considering their individual Brusilovsky [6] distinguished between two different
learning styles. ways of student modelling: collaborative and
automatic. In the collaborative approach, the learners
1. Introduction* provide explicit feedback which can be used to build
and update a student model, such as filling out a
The field of learning styles is complex and affected learning style questionnaire. In the automatic approach,
by several aspects, leading to different concepts and the process of building and updating the student model
views. Many learning style models exist in literature, is done automatically based on the behaviour and
such as the learning style model by Kolb [1], Honey actions of learners while they are using the system for
and Mumford [2], and Felder and Silverman [3]. While learning. The automatic approach is direct and free
there are still many open issues with respect to learning from the problem of inaccurate self-conceptions of
styles, all learning style models agree that learners students. Moreover, it allows students to focus only on
have different ways in which they prefer to learn. learning rather than additionally providing explicit
Furthermore, many educational theorists and feedback about their preferences. In contrast to
researchers consider learning styles as an important learning style questionnaires, an automatic approach
can also be more accurate and less error-prone since it
analyses data from a specific time span rather than data
* The author would like to thank the National Science Council of the which are gathered at one specific point of time.
Republic of China, Taiwan, for financially supporting this research
under Contract No. NSC 097-2811-S-008-001-.

978-0-7695-3167-0/08 $25.00 2008 IEEE 482

DOI 10.1109/ICALT.2008.84
 In this paper, we propose an automatic student discussion forums can give indications about the
modelling approach for identifying learning styles in students preference for active or reflective learning.
learning management systems (LMSs). LMSs such as While active learners are expected to post more often
Moodle [7] and WebCT [8] are commonly and in order to ask, discuss, and explain something,
successfully used in e-learning. They aim at supporting reflective learners are supposed to prefer to participate
teachers in creating and managing online courses and passively by carefully and frequently reading the
provide them with a great variety of features which can postings but only rarely posting by themselves. Due to
be included in the course such as learning material, the preference of testing and trying things out, active
quizzes, discussion forums, assignments, and so on. learners are expected to perform more self-assessment
The proposed student modelling approach is developed tests and more exercises as well as spend overall more
in a generic way, based on commonly used features in time on exercises. Furthermore, they are supposed to
LMSs, and is therefore applicable for LMSs in general. spend only little time on examples since they prefer
Regarding the learning style model, we selected the doing something by themselves rather than looking at
Felder-Silverman learning styles model (FSLSM) [3]. how someone else has solved a problem. Since
FSLSM describes learning styles in very much detail reflective learners like to think and reflect about the
by characterising each learner according to four material, they are expected to visit and spend more
dimensions: active/reflective, sensing/intuitive, visual/ time on reading material like content objects as well as
verbal, and sequential/global. stay longer at outlines. They also tend to take longer on
In the following section, the concept for identifying self-assessment tests as well as on the result pages of
learning styles is introduced. Subsequently, the self-assessments and exercises for reflecting on their
evaluation of the proposed approach and its results are results. As a consequence, reflective learners are also
presented. Section 4 provides discussion about the expected to answer the same question in a self-
proposed approach and introduces related works. assessment test less often twice wrong.
Section 5 concludes the paper.
2.1.2. Sensing/intuitive dimension. Since sensing
2. A concept for identifying learning styles learners favour concrete material like facts and data,
whereas intuitive learners prefer to learn abstract
In the following subsections, the investigated material such as theories and their underlying meaning,
patterns of behaviour for each learning style dimension analysing the performance on questions about facts as
as well as the concept for calculating learning styles well as on theories and concepts provides an indication
from these patterns are presented. about the preferred learning style. Furthermore, in
order to learn from concrete material, sensing learners
2.1. Relevant patterns of behaviour tend to prefer examples. Therefore, the visits and time
spent on examples serve as other patterns. On the other
In order to make our approach applicable for LMSs hand, intuitive learners are supposed to learn from
in general, only commonly used features in LMSs were content objects and use examples only as
selected to be the basis for patterns. These features supplementary material. Therefore, the number and
include: content objects, outlines, examples, self- time spent on content objects tend to be higher and the
assessment tests, exercises, and discussion forums. number and time spent on examples tend to be lower.
Furthermore, the navigation behaviour of students in Furthermore, sensing learners like to solve problems
the course was considered. based on standard procedures, which can be again
In the next subsections, the characteristics of each indicated by a high interest in examples in order to see
learning style with respect to FSLSM are described and and learn existing approaches and a high number of
the relevant patterns for identifying learning styles for conducted self-assessment tests and exercises in order
each dimension are presented, using the literature to check the acquired knowledge. On the other hand,
regarding FSLSM [3] as basis. intuitive learners tend to be more creative and like
challenges. Therefore, they are expected to be better in
2.1.1. Active/reflective dimension. Active learners are answering questions about developing new solutions,
characterised as learners who prefer to process which requires the understanding of underlying
information actively by doing something with the theories and concepts. Another characteristic of
learned material, for example discussing, explaining, or sensing learners is that they are more patient with
testing it. On the other hand, reflective learners prefer details and work carefully but slowly. With respect to
to think about the material and work alone. Regarding the preference for working slowly, the time taken for
discussing and explaining, communication tools like self-assessment tests is considered as pattern. Because
these students tend to check their answers carefully

483
before submitting, another pattern is the number of specific learning style preferences. For example, if a
revisions performed before handing in a test or learner often visited exercises, this gives us a hint that
exercise. Another pattern is the time students spent on the learner prefers an active learning style. Hints are
reviewing their results, where sensing learners again stated by four values: 3 indicates that the students
are expected to spend more time. Furthermore, their behaviour gives a strong indication for the respective
preference for being careful with details can be learning style, 2 indicates that the students behaviour
indicated by their performance on questions about is average and therefore does not provide a specific
details. hint, 1 indicates that the students behaviour is in
disagreement with the respective learning style, and 0
2.1.3. Visual/verbal dimension. While visual learners indicates that no information about the students
learn best from what they can see such as graphics, behaviour is available. In order to classify the
images, and flow charts, verbal learners prefer to learn behaviour of students into these four groups, thresholds
from words, regardless whether they are spoken or from literature are used as basis, considering
written. Therefore, the performance on questions about additionally the characteristics of the respective course.
graphics as well as on text can act as other patterns. By summing up all hints and dividing them by the
Furthermore, verbal learners tend to like number of patterns that include available information,
communicating and discussing with others. Thus, a a measure for the respective learning style is
high number of visits and postings as well as a high calculated. This measure is then normalised on a range
amount of time spent in a discussion forum can from 0 to 1, where 1 represents a strong positive
indicate a verbal learning style. Furthermore, verbal preference and 0 represents a strong negative
learners are expected to visit reading material such as preference for the respective learning style. If no
content objects more often. pattern includes available information, no conclusion
can be drawn.
2.1.4. Sequential/global dimension. Sequential
learners are more comfortable with details, whereas 3. Evaluation
global learners tend to be good in seeing the big
picture and connections to other fields. Therefore, the The proposed student modelling approach was
performance of questions dealing with overviews of evaluated by a course about object oriented modelling,
concepts or connections between concepts and held at a university in Austria. 127 students
questions about details serve as patterns for this participated in the study. The course included all types
dimension. Because global learners are interested in of learning objects described in the previous section.
getting the big picture, outlines of the course and the The LMS Moodle [7] was used and few extensions of
chapters are especially important for them. A high its tracking mechanism were performed, allowing the
number of visits and more time spent on such chapter system to gather data regarding all introduced patterns.
outlines as well as on the course overview page In order to classify the occurrence of behaviour with
indicate a global learning style. Furthermore, their respect to the investigated patterns, thresholds were
interest in relating and connecting topics to each other used for each pattern. In the next subsection, these
helps them to interpret predefined solutions and thresholds are discussed. Subsequently, the evaluation
develop new solutions. Therefore, they are expected to method and the results of the evaluation are described.
perform better on respective questions. The navigation
of learners in a course acts also as a pattern denoting a 3.1. Classifying the occurrence of behaviour
sequential or global learning style. While sequential
learners tend to go through the course step by step in a The thresholds for distinguishing whether students
linear way, global learners tend to learn in large leaps, have a high, moderate, or low occurrence of behaviour
sometimes skipping learning objects and jumping to regarding each pattern were determined from literature
more complex material. Therefore, the number of and adjusted based on the characteristics of the course
skipped learning objects can act as a pattern. rather than depending on the students average
behaviour in the class. Thus, the approach is also
2.2. From behaviour to learning styles applicable for small classes, where the average
distribution of learning style preferences might not
The previous section described the patterns which apply due to its small size.
are incorporated for each dimension as well as whether The thresholds regarding discussion forum were
a high or low occurrence indicates a specific learning based on recommendations from Rovai and Barnum
style preference. Based on this information, data about [9], but were lowered since the forum was mainly used
students behaviour can be used to calculate hints for

484
for asking questions which were then answered by the which was developed by Felder and Soloman [11] for
tutors rather than discussing with other students. identifying learning styles based on the FSLSM. The
Therefore, for the number of visits, thresholds of 7 and proposed approach aims at detecting learning styles for
14 visits per week were used, for the time students each dimension of the FSLSM on a 3-item scale,
spent on the forum, thresholds of 5 and 10 minutes per distinguishing, for example, between an active,
week were used, and for the number of postings, balanced, and reflective learning style. Therefore, the
thresholds of 2 and 4 postings per course were used. measure introduced in Section 2.2 was divided into
Based on the assumptions of Garca et al. [10], the three groups using values of 0.25 and 0.75 as
thresholds for visiting exercises were set to 25% and thresholds. Similarly, results of the ILS questionnaire
75% of available exercises. For self-assessment tests were divided into three groups. For measuring the
and examples, we used a threshold of 50% and 100% precision of the proposed approach, including also how
since both types of learning objects were designed in a close the predicted learning style is to the learning style
way that each object might be visited more than once. based on the ILS questionnaire, the following measure
For outlines, thresholds of 75% and 150% were used. proposed by Garca et al. [10] was used:
Regarding content objects, students had additionally n

the possibility to download the learning material for Sim( LS predicted , LS ILS )
print. Therefore, the content objects were mainly used Precision = i =1
100 ,
for looking up information when students were n
conducting, for example, some exercises or were where LSpredicted refers to the learning style predicted by
reflecting about a topic. Therefore, the thresholds for the proposed approach, using a 3-item scale, LSILS
visiting content objects were set to 10% and 20% of all represents the learning style from the ILS
available content objects. Furthermore, the thresholds questionnaire, using a 3-item scale, and n is the number
for visiting the course overview page was determined of students. The function Sim compares its two
with 10% and 20% of all visited learning objects. parameters LSpredicted and LSILS and returns 1 if both are
The thresholds for the time spent on examples, equal, 0.5 if one represents a balanced learning style
exercises, self-assessment tests, content objects, and the other represents a preference for one of the two
outlines, and the course overview page were poles of the dimension, and 0 if they are opposite.
determined as 50% and 75% in relation to the expected
learning time of students with high interest in the 3.3. Results
respective type of learning object, following the
recommendation of Garca et al. [10]. Table 1 shows the results of the comparison
For the time spent on the results of an exercise or between the proposed approach and the ILS
self-assessment test, thresholds of 30 seconds and 60 questionnaire. The achieved results range from 73.33%
seconds were assumed. Thresholds for the performance to 79.33%, demonstrating a high precision of the
of specific question types were assumed as 50% and proposed approach for all dimensions of the FSLSM,
75% of correctly answered questions, based on the and therefore, show that the proposed approach is
applied grading system. With respect to revisions of suitable for identifying learning styles.
self-assessment tests and exercises, thresholds were
determined as 2.5% and 5% of performed self- Table 1. Results of the comparison
assessment tests or exercises. The thresholds regarding act/ref sen/int vis/ver seq/glo
how often students answered a self-assessment 79.33% 77.33% 76.67% 73.33%
question twice wrong were assumed as 25% and 50%
of times a student is asked the same question twice. 4. Discussion and related work
Regarding skipping learning objects, we looked at
how often students skipped learning objects in relation The proposed approach is based on literature and
to the total number of visited learning objects. applies a simple rule for calculating learning styles
Thresholds of 1% and 2% of times students used the from indications gathered from the students behaviour
navigation menu to skip learning objects were during an online course in an LMS. The underlying
assumed. concept is quite similar to the concept of the ILS
questionnaire, apart that in the automatic approach
3.2. Method of evaluation information from students behaviour is used rather
than asking students about their preferences.
In order to evaluate our approach, its results were Related works aim at identifying learning styles in
compared with the results of the Index of Learning specific learning systems rather than in LMSs in
Styles (ILS) questionnaire, a 44-item questionnaire

485
general. Furthermore, they focus on data-driven students behaviour will be used on the fly for
approaches, where data from students behaviour and modifying and updating the student model and
their learning styles were used in order to build a therefore, allowing the system to immediately respond
model for calculating learning styles. Garca et al. [10] on students needs.
applied Bayesian networks in order to detect learning
styles for three dimensions of the FSLSM. In their final 6. References
study, they achieved results of 58% for the active/
reflective dimension, 77% for the sensing/intuitive [1] D.A. Kolb, Experiential Learning: Experience as the
dimension, and 63% for the sequential/global Source of Learning and Development, Prentice-Hall,
dimension, using the same measure as proposed in Englewood Cliffs, New Jersey, 1984.
Section 3.2. Another study was conducted by Cha et al. [2] P. Honey and A. Mumford, The Manual of Learning
[12], investigating the use of Decision Trees and Styles. Peter Honey, Maidenhead, 1982.
Hidden Markov Models. The results were promising, [3] R.M. Felder and L.K. Silverman, Learning and teaching
styles in engineering education, Engineering Education,
however, only data from the ILS questionnaire 78, 1988, pp. 674681. Preceded by a preface in 2002:
indicating a strong or moderate preference on a http://www.ncsu.edu/felderpublic/Papers/LS-1988.pdf.
specific learning style dimension were considered and [4] N. Bajraktarevic, W. Hall, and P. Fullick,
data indicating a balanced learning style were Incorporating Learning Styles in Hypermedia
excluded. Therefore, further investigations towards a Environment: Empirical Evaluation, Proceedings of the
more accurate approach are necessary. Workshop on Adaptive Hypermedia and Adaptive Web-
Another important issue for the automatic Based Systems, Nottingham, UK, 2003, pp. 41-52.
identification of learning styles is the considered [5] S. Graf and Kinshuk, Providing Adaptive Courses in
number of patterns per learning style dimension. For Learning Management Systems with Respect to
Learning Styles, Proceedings of the World Conference
the proposed approach, each learning style dimension on E-Learning in Corporate, Government, Healthcare,
consists of a relatively high number of patterns, and Higher Education (eLearn), AACE, 2007, pp.
compared to those of related works, such as the model 2576-2583.
introduced by Garca et al. [10] as well as one of our [6] P. Brusilovsky, Methods and techniques of adaptive
previous research work [13]. On one hand, a high hypermedia, User Modeling and User-Adapted
number of patterns give more detailed information and, Interaction, 6, 1996, pp. 87-129.
on the other hand, it helps to identify learning styles in [7] Moodle, 2008. Available at http://www.moodle.org.
LMSs in general since information about some patterns Accessed 8 January, 2008.
might not be available in all LMSs. [8] WebCT, 2008. Available at http://www.webct.com.
Accessed 8 January, 2008.
[9] A.P. Rovai and K.T. Barnum, On-line course
5. Conclusions effectiveness: an analysis of student interactions and
perceptions of learning, Journal of Distance Education,
This paper introduced an automatic student 18 (1), 2003, pp. 57-73.
modelling approach for identifying learning styles [10] P. Garca, A. Amandi, S. Schiaffino, and M. Campo,
based on the FSLSM in LMSs. The proposed approach Evaluating Bayesian Networks' Precision for Detecting
uses the behaviour of students during they are learning Students' Learning Styles, Computers & Education, 49
in order to gather hints about their learning styles. By (3), Elsevier, 2007, pp. 794-808.
[11] R.M. Felder and B.A. Soloman. Index of Learning
applying a simple rule-based mechanism, learning
Styles Questionnaire. Retrieved 6 January, 2008, from
styles are calculated based on the gathered indications. http://www.engr.ncsu.edu/learningstyles/ ilsweb.html
The evaluation of the approach demonstrated good [12] H.J. Cha, Y.S. Kim, S.H. Park, T.B. Yoon, Y.M. Jung,
results and showed that the approach is suitable for and J.-H. Lee, Learning Style Diagnosis Based on User
identifying learning styles with respect to the FSLSM. Interface Behavior for the Customization of Learning
Since the proposed approach is applicable for LMSs Interfaces in an Intelligent Tutoring System,
in general rather than for one specific system, it allows Proceedings of the 8th International Conference on
teachers to identifying their students learning styles Intelligent Tutoring Systems, Lecture Notes in Computer
while they hold their courses in LMSs. The Science, Berlin, Heidelberg, Springer, Vol. 4053, 2006,
pp. 513-524.
information about students learning styles can be used
[13] S. Graf and Kinshuk, An Approach for Detecting
for making students aware of their learning styles and Learning Styles in Learning Management Systems,
providing them with courses/material that fit their Proceedings of the International Conference on
learning styles. Advanced Learning Technologies (ICALT), Los
Future work will deal with developing a concept for Alamitos, CA, IEEE Computer Science, 2006, pp. 161-
dynamic automatic student modelling, where data from 163.

486