Combining Learner’s Preference and Similar Peers’ Experience in
Adaptive Learning
Fandry Indrayadi and Dade Nurjanah
Telkom University, Jl. Telekomunikasi 1, Bandung, Indonesia
Keywords: Adaptive Educational Hypermedia, Domain Model, Cognitively-Oriented Method, Open Learner Model,
Learner Preference Pattern.
Abstract: Adaptive educational hypermedia (AEH) offers learning adaptation and personalization. In terms of
adaptation, AEH plays the role of a tutor and controls learning. To the contrary, personalization gives
learners the freedom to explore the materials they consider necessary. Challenges emerge in respect of
improving adaptation and preventing learners from getting lost when exploring concepts and materials in
the large. This paper discusses approaches to improve adaptation and personalization. A knowledge map
that organizes and visualizes the domain model has been developed using a cognitively-oriented method. It
combines the individual learner’s progress and preference with similar peer experiences to improve
adaptation. Furthermore, it implements an open learner model to nurture self-progress awareness.
1 INTRODUCTION
Current advancements of technology have made
indirect learning viable. This changes how the
learning process goes, such as removing the
requirement for learners and teachers to be in the
same place. Web-based courses are one of the many
media that can be used for indirect learning.
However, while many research papers and media
publications report substantial success with Web-
based education, a careful analysis of the situation
and informal discussions with "on-line teachers"
show that Web-based education is quite far from
achieving its main goal. In many current Web-based
courses, the course material is still implicitly
oriented to a traditional on-campus audience, which
means reasonably homogeneous, reasonably well-
prepared and well-motivated students who have
access to teachers and assistants to fill possible gaps
and resolve misunderstandings (Oneto et al., 2009).
A web-based education should be aimed at a larger
audience with different knowledge, goals and
learning capabilities.
That is where Adaptive Educational Hypermedia
(AEH) comes in. AEH is a system that can adapt to
the learner, helping tutors to create a learning
process that is relevant to the learner's needs,
recommending learning objects and enabling
learners to choose what they like. Such a system is
also capable of helping learners in their self-
assessment and the personalisation of the learning
process.
A challenge occurs regarding how to improve
adaptation so that the recommended artefacts suit
learners’ needs. Sometimes a learner model is not
enough, for instance when learning has just started
and the learner model does not contain much
information about the learner. Many adaptive
systems provide a default learning scenario to
anticipate this situation and it works. However, a
default scenario contradicts the principle of adaptive
learning.
A potential solution comes from a recommender
system. Since the adaptation model of AEH is like a
recommender system, the process of recommending
something by referring to a similar case can be
adopted in AEH. This idea meets the principle of
social learning that a learner learns better when
he/she is learning with experienced learners
(McLeod, 2007; Vygotsky, 1978). A question then
emerges regarding which experienced learners will
help a learner to get the most suitable learning
materials.
Another challenge occurs regarding the
accessibility of the learner model. In web-based
education, it is common that only teachers, not
learners, can see the learner model. It is also
common that learning objects are presented
486
Indrayadi F. and Nurjanah D..
Combining Learner’s Preference and Similar Peers’ Experience in Adaptive Learning.
DOI: 10.5220/0005490904860493
In Proceedings of the 7th International Conference on Computer Supported Education (CSEDU-2015), pages 486-493
ISBN: 978-989-758-107-6
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
pedagogically, disregarding whether they are
relevant to the learner's needs or not. To build a
system able to support self-assessment and learning
process personalisation efficiently, we need both a
user interface that will represent learner models
intuitively and a recommender that is able to direct
the right learning object to the right learner.
This paper addresses two questions:
• First, from the perspectives of adaptive learning
and collaborative learning, how to combine an
individual user model with peers’ experience;
• Second, what features are necessary to support
adaptation and personalisation, thus making
learners aware of their progress and able to
understand the path to achieve their learning
goals.
To solve the problems, we propose the
integration of the Open Learner Model (Bull and
Kay, 2010) and the Learner Preference
Pattern(Wang et al., 2007). Open Learner Model
functions as a visual interface guideline, while
Learner Preference Pattern produces
recommendations of suitable materials for learners.
The domain model is implemented with a
cognitively-oriented method (Liang et al., 2012), in
the form of a knowledge graph.
2 RELATED WORK
2.1 Adaptive Educational Hypermedia
Adaptive Educational Hypermedia (AEH), known as
a kind of adaptive learning system, builds a learning
model based on the knowledge, preferences and
goals of the learner. Unlike conventional e-learning
where learners have the same learning object on the
same course, this system can adapt and recommend
a relevant learning object. As a learner's needs,
preferences and goals change, the AEH should
always oversee these changes to update the learner
model. In general, the framework of AEH can be
illustrated below:
Figure 1: Diagram of AEH Framework (Triantafillou et
al., 2003).
2.2 Open Learner Model (OLM)
Open Learner Models are learner models that can be
viewed or accessed in some way by the learner, or
by other users, such as teachers, peers or parents.
Their goals are to visualise knowledge, preferences
and cognitive skills intuitively. This can be done
using an interface designed for the learner or, in
some cases, other people that will help the learning
process.
OLM aims to be helpful to the learner as
identified in the SMILI (Student Model that Invite
the Learner In) OLM Framework as (Brusilovsky et
al., 2011; Bull and Kay, 2010):
• Enabling metacognitive activities, such as
planning and self-monitoring;
• Giving learners greater control and responsibility
in learning processes.
• Supporting collaborative learning;
• Helping learners to interact well with peers,
teachers and parents;
• Providing navigation to suitable materials,
exercises, problems, activities or tasks;
• Supporting formative and summative
assessments.
A former study that implements OLM is
QuizMap (Brusilovsky et al., 2011). It implements a
pedagogically-oriented knowledge model with OLM
that enables learners to know their progress and
which questions they can choose .OLM provides
several concepts that can be implemented on an
effective interface, such as OLMlet that integrates
cognitively-oriented knowledge space with a learner
model(Bull and Kay, 2010).
2.3 Learner Preference Pattern
Personalised recommendation mechanisms which
take into account peers’ experience have been
proposed in a number of previous study (Troussas et
al., 2013; Wang et al., 2007). The method proposed
by Tzone I Wang et al. (Wang et al., 2007)
implements two algorithms, the preference-based
algorithm and the correlation-based algorithm, to
rank the recommended results to advise a learner
concerning the most suitable learning objects. This
model uses a specific ontology of a certain course to
infer which learning objects are needed for a learner.
The inference is based on his/her past studying
histories that are recorded as the learner’s personal
preference pattern.
Another consideration in selecting learning
objects is by referring to the experience of similar
learners (Wang et al., 2007). The similarity of
CombiningLearner'sPreferenceandSimilarPeers'ExperienceinAdaptiveLearning
487
learners can be inferred from similar values of
certain parameters
3 OUR RESEARCH
In response to the aforementioned research
questions, we have implemented an AEH with
combined user model and peers’ experience and
applied OLM. The system is divided into 3 parts,
illustrated as follow:
Figure 2: Overall System Structure.
The first part (illustrated by the left part) is the
interface of system and learners, implemented
visually based on cognitively-oriented modelling. In
this part, the learner interacts with the system
directly, such as by doing an assessment via pre-test,
giving feedback, or reading topological maps. The
second part (illustrated by the middle part) is the
result of the learner and the system's interaction,
contained in learner models. In our case, learner
models are a sequential file stored in the hard disk.
Finally, the third part (illustrated by the right part) is
data storage for the course learning object,
knowledge/domain model and learner model.
3.1 Domain Model
Adaptive learning must be supported by a large
networked knowledge space and a huge volume of
learning materials in various formats. There are two
main approaches to the model domain model of
AEH; namely, the pedagogically-oriented topic
based modelling, which is a taxonomy of coarse-
grain topics that uses a tree as a structure
(Brusilovsky et al., 2011; Sosnovsky and
Brusilovsky, 2005), and the cognitively-oriented
concept based modelling, which is a link of fine-
grain concepts that uses a graph as the visualisation
(Brusilovsky et al., 2011). The pedagogically-
oriented method is commonly applied as it provides
a firm hierarchical structure among topics that can
support a sequential flow of learning. This method
however, has a disadvantage in that, if the hierarchy
is too deep, it may result in boredom for students as
they must complete all the materials at the deepest
level before they can progress to another topic.
On the other hand, the cognitively-oriented
method organises a curriculum in the form of a
concept graph. This method is suitable for allowing
students to explore learning material concepts
without restraint. The idea is that students learn, gain
understanding and create connections or associations
between the concepts. The disadvantage of this
method is that students may get lost as they learn
many topics at random and it may, therefore, result
in students failing to attain the learning objectives.
Figure 3: Cognitively-oriented domain model of Data
Structure.
We have implemented cognitively-oriented
domain model for Data Structures, one subject
taught in the Computer Science undergraduate
programme. In general, domain knowledge attributes
consist of:
• Concept's ID and label
• Connection between concepts
• Learning objects related to the concept,
containing ID, external link towards said learning
object, and tags/keywords related to said learning
object.
3.2 Learner Model
In this research, we combine individual and social
CSEDU2015-7thInternationalConferenceonComputerSupportedEducation
488
learning to perform adaptive navigation. Social
learning is implemented in the form of collaborative
tagging. The learner model records a learner’s
progress, tags and ratings on learning materials. To
navigate, firstly, learners have to take a competence
test. As a result, learners’ competence on each
concept can be assessed. Finally, a list of learning
materials to suit the learners’ competence is
delivered.
The competence test consists of several questions
with each question related to each concept in the
course. Students’ answers are categorised as correct,
incorrect or cannot be justified. Correct/incorrect
answers will categorise the learner as a student
having expert/intermediate cognitive skills on
corresponding topics. On the other hand, when the
learner answers “I do not know” for a question,
he/she will be categorised as beginner on the
corresponding topic. The cognitive skills will affect
which learning objects to be recommended to the
learner. To have more accurate learner profiles,
learners can override their profiles by conducting a
self-assessment.
Figure 4: Competence test.
Cognitive skills are not the only information
recorded in the learner model. The learner model
contains cognitive, preferences and the learning
progress of a learner, as well as the tagging and
rating he/she put to learning objects. Every time a
learner has learned a new material, she/he will asked
to give a feedback in the form of rating:
• 1, if the material is not helpful at all.
• 2, if the material is not really helpful.
• 3, if the material is quite helpful.
• 4, if the material is really helpful.
The rating will be recorded in the learner’s model
and will be used in the recommendation process for
her/himself or other learners who have similar
models. Table 1 presents the attributes of the learner
model.
Table 1: Domain knowledge attributes.
Attributes Type Description
Id String id of learner
RatingObjects Array
<String>
Array of objects current
learner have given feedback
RatingValues Array
<Float>
Values of objects current
learner have given feedback
Tags Array
<String>
Array of tags/keywords this
learner is probably
interested in
TagValues Array
<Float>
Values of preference score
for each tag
CognitiveObjects Array
<String>
Array of concepts this
learner has learned
CognitiveValues Array
<Int>
Values of cognitive skills
this learner currently has
("beginner", "intermediate",
or "expert")
On visualising the learner model on the
topological map, the system implements an overlay
model combined with OLM. It uses colour codes to
intuitively label nodes with the corresponding
learner's cognitive skill, grey for beginner when the
topic is not learned yet, red for intermediate when
the topic is being learned, and green for expert when
the topic has been mastered. Implementing OLMlet's
skill meter (Bull and Kay, 2010)and structured view
as a guideline, we can visualise colour-coded learner
models on a topological map as shown in Figure 5.
The progress bar indicates the learner's learning
progress on each concept; it is empty when the
learner has not learned anything and fills up each
time the learner reads some learning material related
to each corresponding concept. Each node in the
graph can be clicked to open the detail panel of the
corresponding concept. It contains a definition, the
learner's cognitive skill, as well as learning
materials. Learning objects on the detail panels are
sorted based on their recommendation score.
Learning objects with the highest RS are then
highlighted and marked as recommended to read.
CombiningLearner'sPreferenceandSimilarPeers'ExperienceinAdaptiveLearning
489
Figure 5: Competence test.
A challenge in producing adaptive learning
based on the learner model and peers’ experience is
to find peers who have a similarity with the learner
and have mastered the topic learned. The experience
of similar learners is considered more useful than
that of the other learners. Similarity can be inferred
from any attribute in the learner model. Our research
refers to the tags and ratings they gave to the
learning materials.
Adaptation is performed in the form of adaptive
navigation as shown in Figure 6.
Figure 6: Learner model life cycle.
Learning objects on the detail panels are sorted
based on their recommendation score. Learning
objects with the highest RS are then highlighted and
marked as "recommended" to read.
3.3 Learning Object Recommendation
The chart below illustrates the calculation process of
calculating recommendation score:
Figure 7: Learning object recommendation.
We improved the Learner Preference Pattern
method (Wang et al., 2007) to calculate the
recommendation score. Each learning object in a
concept is sorted based on their RS from the highest
to the lowest one. There are 2 factors affecting the
recommendation score: the preference score that
represents which materials interest a learner to learn
and the helpfulness score that formulates which
materials the learner considers necessary. We
CSEDU2015-7thInternationalConferenceonComputerSupportedEducation
490
consider that the preference score and helpfulness
score equally influence the recommendation score.
Hence we define the recommendation score for a
learning object as follows:
0.5 * preference score + 0.5 * helpfulness score
To calculate the preference score, first of all we need
to calculate the Basic Preference Weight (BPW), a
weight that represents the degree of a learner's
preference for a feature value in a feature [6]. In our
case, the features are the tags each learning object
has. These tags vary. For example, "video" for
learning object links using video as a learning
medium or "english" if the object uses English as the
language. To get BPW, the value of a tag on a
learning object, the score of such a tag is divided by
the maximum score of the learning object tags.
BPW of the k-th feature value of the i-th feature =
(The preference score of the k-th feature value) / (the
maximum preference score of tags given to the
learning object)
The preference score for a learning object is
calculated by summarising all the BPW scores of
tags given to the learning objects, and then dividing
by the number of tags given to the learning objects.
In addition to the preference score, to increase
the accuracy of recommendations, feedback from
other learners with similar experience and
preferences is taken into account. The similarity of
two learners, called learner1 and learner2, is
calculated as follows:
Sim(learner1, learner2)=
Sum ((rating1-avg1)(rating2-avg2)) /
Sqrt(Sum ((rating1-avg1)
2
(rating2-avg2)
2
))
Where rating1 and rating2 are feedback scores that
learners 1 and 2 have given to learning objects; avg1
and avg2 are the averages of feedback scores they
have given. The formula is applied to learning
objects they have both learned.
The system will first iterate all the learner profile
database and calculate each one's similarity. A
perfectly-similar learner compared to a currently
active learner will have a similarity of 1. A learner
having similarity of more than 50% (0.5) will be
considered "similar enough” and will be included in
a group of similar learners. Using the similar
learners group, we can then calculate the helpfulness
score of a learning object for a learner by the
following formula:
The average of all feedback given by the learner +
Difference_error_score.
The Difference error score is taken into account to
counterbalance the difference between a learner and
the similar learners’ consideration in rating a same
learning object (lo). The difference error score
summarises the difference between the learner and
each similar learner as follows:
Difference_Error_Score =
Sum (difference_score (L_id, L_simm)
Where (difference_score (L_id, L_simm) is equal to:
(ratings(L_simm, lo) – avg(ratings(L_simm))
*Sim (L_id, L_Simm)
After accessing the recommended learning
objects, learners can give a feedback rating in the
scale of 1 to 4 (very helpful, quite helpful, not really
helpful, not helpful at all). Learner's profile, Tags
and TagValues representing the learner's preference
toward a certain tag will be updated. The change in
the Learner's profile follows these rules:
• a feedback rating of 1 or 2 will not change the
learner's profile
• a feedback rating of 3 or 4 will increase all
preference scores of tags in the learning object by
Δs amount, as expressed by the following:
(feedbackRating * totalTags) /
(4 * the number of learning objects having the
same tag)
4 TESTING
We have conducted two kinds of test. The first test is
a comparison test between two cases, to test whether
the tool takes into account learner models in
recommending suitable learning materials. The
correctness is detected from different
recommendations that should be produces for
different learners. The second test isa usability test
by eliciting learners’ experience of using this AEH.
There are three parameters tested including
learnability, helpfulness and efficiency. This paper
focuses on the first test.
In the first test, a comparison of
recommendations for two learners having different
preferences was carried out. The first learner
considers Indonesian articles are helpful and the
second learner considers that watching English
videos is helpful for his learning. To find out the
CombiningLearner'sPreferenceandSimilarPeers'ExperienceinAdaptiveLearning
491
learners’ skills, they were required to complete the
pre-test. It resulted in the learner models shown in
Figures 8 and 9. They presented different cognitive
skills of the two learners. The first one, as shown in
Figure 8, is novice in all topics, except in the first
and second topics. The second learner, as shown in
Figure 9, has expertise in many topics. A
comparison of learning material recommendations is
applied to the Pointer topic, which has not been
learned by both learners.
Figure 8: A model of a learner who prefers Indonesian
articles.
Figure 9: A model of a learner who prefers English
materials.
The process to find recommended materials for
the first learner will find materials which are articles
written in Indonesian and for beginners and it is
based on similar learners’ experience. In this case,
two learners were detected having similar models.
The recommended materials are sorted and
highlighted.
Figure 10: A learning object recommendation for the first
learner.
On the other hand, the recommendation for the
second learner presents different learning materials.
Figure 11: A learning object recommendation for the
second learner.
5 CONCLUSIONS
In this paper, an implementation of the Open
Learner Model and Recommendation based on
Learning Preference Pattern in an Adaptive
Educational Hypermedia is proposed to help self-
learning for Data Structure. A cognitively-oriented
Open Learner Model provides a guideline for an
intuitive cognitively-oriented model suited to these
needs. Colour codes for visualising learners’
cognitive skills and a progress bar to track the
learner's learning progress are also applied. Learning
Objects Recommendation based on the Learner
Preference Pattern are capable of recognising adapt
on feedback and the changes learners made, then
gives them learning object recommendations tailored
to their current preference and cognitive skills.
CSEDU2015-7thInternationalConferenceonComputerSupportedEducation
492
In a real case of AEH, OLM and learner
preference pattern are correlated. OLM is applied as
a guideline for a domain model implemented with a
cognitively-oriented method that represents domain
knowledge and materials in a graph of concepts.
Hence, it could help learners for self-assessment. On
the other hand, the learner preference pattern is
applied for finding and recommending learning
objects. It controls learners by recording their
cognitive skill progress and preference, and then
adapts their learning based on the learner model.
As our research is tested with a small number of
participants and limited preferences, further
development of this research can include more
preferences and applied to larger participants.
Furthermore, the future work can be focusing on
learner similarity. As learner model captures various
parameters of learners, there might be many
combinations of parameters to be considered in
detecting learner similarity.
REFERENCES
Brusilovsky, P., Hsiao, I.-H. & Folajimi, Y. (2011).
QuizMap: open social student modeling and adaptive
navigation support with TreeMaps. Proceedings of the
6th European conference on Technology enhanced
learning: towards ubiquitous learning. Palermo, Italy.
Bull, S. & Kay, J. (2010). Open Learner Models. In
Advances in Intelligent Tutoring Systems Studies in
Computational Intelligence, pp. 301-322. Springer.
Liang, M., Guerra, J. & Brusilovsky, P. (2012). Building
multi-layer social knowledge maps with Google Maps
API. Workshop on Semantic and Adaptive Social Web
(SASWeb 2012). Montreal, Canada.
Mcleod, S. (2007). Vygotsky.
From:http://www.simplypsychology.org/vygotsky.htm
l. Accessed on 16 June, 2012.
Oneto, L., Herder, E., Abel, F. & Smits, D. (2009).
Making today's Adaptive Learning Systems adaptive.
In the proceeding of EC-TEL 2009, pp. Springer
Berlin/Heidelberg.
Sosnovsky, S. & Brusilovsky, P. (2005). Layered
Evaluation of Topic-Based Adaptation to Student
Knowledge. Fourth workshop on Evaluation of
Adaptive Systems at the 10th International User
Model Conference.
Triantafillou, E., Pomportsis, A. & Demetriadis, S. (2003).
The design of the formative evaluation of an adaptive
educational system based on cognitive styles.
Computers & Education, vol. pp. 87-103.
Troussas, C., Virvou, M. & Alepis, E. (2013). Comulang:
towards a collaborative e-learning system that
supports student group modeling. SpringerPlus.
http://www. springerplus.com/content/2/1/387.
Vygotsky, L. (1978). Mind in Society. 1st edition. USA:
Harvard University Press. As cited by Mcleod, Saud.
(2007). Vygotsky. http://www.simplypsychology.
org/vygotsky.html.
Wang, T. I., Tsai, K. H., Lee, M. C. & Chiu, T. K. (2007).
Personalized Learning Objects Recommendation
Based on the Semantic-Aware Discovery and the
Learner Preference Pattern. Educational Technology &
Society, vol. 10, pp. 84-105.
CombiningLearner'sPreferenceandSimilarPeers'ExperienceinAdaptiveLearning
493
