Leveraging User Profile Attributes For Improving Pedagogical Accuracy of Learning Pathways

Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Leveraging user profile attributes for improving

pedagogical accuracy of learning pathways


Tanmay Sinha Ankit Banka Dae Ki Kang*
Department of Computer Science Department of Computer Science Department of Computer Engineering
Vellore Institute of Technology Vellore Institute of Technology Dongseo University
Chennai 600127, India Chennai 600127, India Busan 617716, Korea
tanmay.sinha655@gmail.com ankitbanka17@gmail.com dkkang@dongseo.ac.kr

like Khan Academy, Merlot(with more than 20,000 learning


Abstract - In recent years, with the enormous explosion of
web based learning resources, personalization has become a resources and about 70,000 registered users), OER
critical factor for the success of services that wish to leverage the Commons(with about 18,000 resources), European Schoolnet's
power of Web 2.0. However, the relevance, significance and Learning Resource Exchange(with more than 43,000 learning
impact of tailored content delivery in the learning domain is still resources from 25 different content providers in Europe and
questionable. Apart from considering only interaction based beyond), online education platforms like Coursera, Udacity,
features like ratings and inferring learner preferences from them, EdX, Khan Academy, Venture Lab and e-learning platforms
if these services were to incorporate innate user profile attributes NPTEL have been established. They provide personalized
which affect learning activities, the quality of recommendations learning environments and offer pedagogical support for
produced could be vastly improved. Recognizing the crucial role
learners in informal educational settings.
of effective guidance in informal educational settings, we provide
a principled way of utilizing multiple sources of information from Apart from providing custom self paced learning tools,
the user profile itself for the recommendation task. We explore such organizations also measure student progress and provide
factors that affect the choice of learning resources and explain in
targeted interventions. In contrast to formal education systems
what way are they helpful to improve the pedagogical accuracy of
learning objects recommended. Through a systematical
in schools and universities where predefined course structures,
application of machine learning techniques, we further provide a well formed accreditation procedures and a regular quality
technological solution to convert these indirectly mapped learner check of knowledge inflow and outflow is maintained by
specific attributes into a direct mapping with the learning domain experts, informal learning is basically a kind of
resources. This mapping has a distinct advantage of tagging lifelong learning activity. Here, students are responsible for
learning resources to make their metadata more informative. The their own learning pace and path [1]. The process of learning
results of our empirical study depict the similarity of nominal is often self directed and heavily dependent on users
learning attributes with respect to each other. We further succeed preferences. Because of no directed course plan and massive
in capturing the learner subset, whose preferences are most likely
enrollment to form such informal learning networks, these
to be an indication of learning resource usage. Our novel system
filters learner profile attributes to discover a tag that links them
learners are often overwhelmed with a plethora of alternative
with learning resources. learning resources to choose from. It is highly impossible for a
faculty to cater to the individual problems of learners. As
Keywords - Apriori principle, Clustering, NMF, Learning pointed out in [25], the problem of improving the quality of
Pathway, Pedagogical accuracy, Technology enhanced learning. education is quite multifaceted and intricate. However, the
suggestion of correct pathways to be followed while studying
I. INTRODUCTION
important concepts, can benefit these learners greatly. Though
Education has been acknowledged as one of the major the field of TEL Recommender system(TEL RecSys) [2]
pillars to support and improve lives of the economically addresses this scenario of mass customization, the major factor
disadvantaged [26]. Technology enhanced learning(TEL) aims for deciding the most appropriate learning activities is still
to enhance learning practices of individuals and organizations learner's rating, review of resources and his fundamental
by the design, development and testing of interactive socio demographic information. This may work well for e-commerce
technical innovations [2]. It provides an opportunity to unleash recommender systems, but similar applications in the domain
the full potential of machine learning algorithms and apply of learning may not be suitable.
them to quantify and incorporate psychological factors that
Basically there are the two major differences between
affect learning. As the web is becoming more ubiquitous, an
product recommendations and TEL RecSys that need to be
increased number of learners are gradually turning towards
considered. Firstly, learning is an activity that takes more time,
online education to seek additional support and guidance.
considering the different study goals of individuals and their
Therefore, improving learner experience is also another major
grasping capabilities. Also, the end objective may be variable.
focus of TEL. Over the years, various learning repositories

*
Corresponding author: Tel.: +82513201724, Email: dkkang@dongseo.ac.kr
Learners may wish to have different accomplishments after them[7].
completion of a learning resource. Secondly, the most 5)K-means clustering- Intuitively, what clustering techniques
preferred(rated) learning path may not be pedagogically the aim to achieve is finding data points which are more similar to
most adequate [3]. Thus, the educational specific needs and one another than the rest of the points in n-dimensional feature
information seeking tasks of stakeholders, make the design of space. K-means is a widely adopted clustering method that
learning resource recommendation quite complex [16]. starts out by creating 'k' different clusters and assigning each
data point to one of the clusters based on some standard
Our statement of contribution in this paper is to develop a
distance measure like Euclidean, Pearson correlation etc. At
system that can effectively categorize resources based on
every iteration, centroid of the clusters are updated with the
inherent factors that affect student's learning. We aim to
mean value of the data points assigned to that cluster. The idea
extract such learning domain specific attributes of learners,
is to minimize intra cluster distance and maximize inter cluster
that have been discovered from past theoretical research. Such
distance. Proposed in [22], K means clustering has been a
critical factors belong to an entirely different context and play
widely studied unsupervised machine learning algorithm.
a much more significant role, as compared to basic profile
6)Apriori principle- Originally presented in [23], it optimizes
information like name, age, sex and other demographic factors.
association analysis in large scale machine learning. Either for
We believe that our system has the potential of improving
finding out frequent itemsets or discovering association rules
alignment of learner's preferences with the learning resources
in large datasets, apriori principle is applied to reduce the
they are using. Also, it could increase learning efficiency,
number of possible interesting itemsets. Measures like support
through a more structured guidance that is well customized
and confidence levels are established to evaluate the
and adapted to the learner's objectives.
associations discovered. Apriori principle says that if an
This paper is structured as follows: In section II.I, the itemset is infrequent, all it's supersets will also be infrequent.
technical terms used are defined. Then, in the following At every iteration of the algorithm, candidate itemsets are
sections II.II and II.III, the user profile attributes are described pruned if their support levels are below the minimum support.
and challenges are presented. The hypothesis, problem The process of joining lower order itemsets based on this
statement and the proposed method is described in detail in criteria is repeated, until no more itemsets can be created
section III. Section IV talks about the experimental results and further.
discussions. Prior work is discussed in section V. The 7)Non negative matrix factorization(NMF)- Formally
conclusion and future work follow at the end in section VI and defined in [24] for learning parts of objects and combining
VII. them to get an aggregate representation, NMF derives it's roots
from linear algebra. The basic idea of NMF is that it
decomposes matrix A into two matrices B(features) and
II.I. PRELIMINARIES
C(weights) and also enforces a non negative constraint on the
entries of these matrices. In machine learning, it is used as a
1)Learning Management system(LMS)- It is a system that
feature extraction algorithm, where the decomposed matrices
assists learners in finding learning resources and guides them
represent the extracted features in the observations and their
to make the best choices in managing these resources. It also
importance or relevance to the observed data.
supports these learners in achieving their goals in a specific
domain. A good literature review of LMS can be found in [4].
2)Learning pathway- A set of learning objects like books,
multimedia(audio recordings, videos), images, slides etc which
II.II. USER PROFILE ATTRIBUTES
are packaged and organized in a sequence form a learning
pathway. A formal definition of the same can be found in [5]. User model is a distinctive feature of adaptive systems, as this
3)Transfer Learning- The process of applying concepts and information allows the system to behave differently for
skills learned in one domain to another related or disjoint different users [20]. In our work, we consider the following
domain in different scenarios is called transfer learning [6]. user profile attributes for incorporating into the proposed
The absence of a guide to assist learners in providing a good machine learning framework to tag learning resources. The
cross domain application context, makes an understanding of objective is to improve the pedagogical accuracy and
the research issues associated with transfer learning quite adequacy of TEL RecSys, by suggesting resources that reflect
difficult. the learner's objective and form a close match to his learning
4)Pedagogical adequacy(accuracy)- It is the kind of requirements. By doing so, the learner can be kept motivated
education that focuses equally on five different aspects of to complete his learning activities in an effective and efficient
learning- student centred tutelage with an understanding of his manner.
educational background, strategic course instruction design to
maximize learning, evaluation of student's advancement 1)Skill level- This includes current competency level of the
towards the course learning objectives for knowledge learner and the target skill level that is to be achieved. Based
retention, use of different teaching strategies, awareness on Bloom's taxonomy and CDIO report[8], the following
regarding teaching challenges and knowing ways to handle proficiency levels are identified:
 Level 1- To have experienced or been exposed 4)Preferred Presentation style- Because of the difference in
to(Knowledge)- refers to the process of recalling cognitive processing , we consider visual, textual and auditory
primer knowledge on a subject presentation styles. More specifically, we include:
 Level 2- To be able to participate in and contribute  Portable document format(pdf) documents
to(Comprehension)- refers to the ability to describe  Power points
and define what is learnt  Web pages(html)
 Level 3- To be skilled in the practice or  Videos
implementation of(Application)- refers to the process  Audio books and resources
of execution of concepts An alternative to the above styles include presentation of
 Level 4- To be able to understand and examples, presentation of theoretical knowledge and practical
explain(Analysis)- refers to the process of analysis, examples [10]. Over and above all these explicit factors that
interpretation and discussion on results can be gathered by an initial questionnaire, other implicit
 Level 5- To be able to lead or innovate in(Synthesis)- factors like effort of the learner and his willingness to obey
refers to the process of creation of thoughts and and follow the recommendations can be measured during
development of new ideas course progress [11][12].
 Level 6- To be able to assess and evaluate(Evaluate)-
refers to the process of estimation and evaluation of
the concepts learnt II.III. CHALLENGES AND MOTIVATION
2)Preferred Learning strategy- This basically refers to the
Though technology alone cannot solve complex problems like
intended way of understanding information, solving problems,
checking and evaluating the knowledge gained. Based on the providing universal and high quality education, but it can
strategic instruction model proposed by [9], we have included surely improve the quality of solutions and their impact on
the following five learning strategies for the purpose of our learning. This is the driving factor behind our work.
study:
 Assignment Completion Strategy- designed to enable 1)We notice that a major issue in TEL RecSys is the use of
students to complete and hand in the assignments on only learner ratings, reviews and basic profile information for
time providing tailored learning pathways. Other critical factors
related to student's academic profile and interests are left out.
 Test-Taking Strategy- designed to be used while
This leads to inaccurate and unsatisfactory recommendations
taking classroom(physical/virtual) tests, with students most of the times. However, because learning takes place in
allocating time and certain priority to each section of extremely diverse and rich environments, we include the other
the test dimensions as stated in Section II.I to make the learning model
 Mnemonic Strategy- designed to help students more helpful and close to real life behaviour. Because these
remember complex concepts using simple user specific attributes do not have a direct link with the
representations(diagrams) and mnemonic codes. learning resources and due to the unavailability of tagged data
for these resources, there is a need to incorporate such
 Self-Questioning Strategy- designed to help students parameters into the recommendation framework. And our
create their own motivation while reading by creating proposed solution finds a way to fill this missing link by using
questions in mind, predicting answers, searching, machine learning algorithms to accomplish the task at hand.
verifying and paraphrasing them
 Error Monitoring Strategy- designed to help students 2)The other challenge while applying machine learning to
independently detect and correct errors make appropriate sense of the user profile attributes is that,
 Teamwork Strategy- designed to provide a framework though skill levels and learning time can be inferred
for organizing and completing specific tasks in small numerically, learning strategy and presentation style are very
groups nominal in nature. Deriving similarities between them is quite
3) Available Learning time - This includes the number of difficult numerically. We cannot represent such nominal
time units(hours per day and number of weeks), and indicates variables on a continuous scale and perform arithmetic
the time allotted for the purpose of study in an informal operations on them. For example, it's difficult to comprehend,
educational setting. how similar or dissimilar test taking and mnemonic learning
strategies are, without any training data. This fact has useful
implications, because distance measures like Euclidean,
Pearson correlation etc. can't be applied to address such a III.II. SOLUTION & APPROACH FOLLOWED
situation. Thus, there is a need to characterize the data further
to leverage these nominal attributes. Step 1- Alg. "build_learner_subset"
input- learner set
output- learner subset based on rating threshold criteria
a. l  L build V  U | riy   0 ,  0  10 and  0  N (i=1
y

III. METHODOLOGY to n, y=1 to m) /*build learner subset*/


b. ui V , consider pij as a point in 5-D feature space (i= 1
Our methodology for tagging learning resources consists of
to n, j= 1 to 5)
deriving a subset of learners who have rated the learning
resource above a particular threshold, finding a relation Fig (i) Alg. "build_learner_subset"
between their nominal attribute parameters, grouping these
learners to find the cluster of largest size, and using the most In Fig (i), for all learning resources, we get a list of learners
frequently occurring learning profile attributes as a more who have rated them greater than a threshold value (e.g.-
informative tag for the learning resource. greater than 9/10). This list is an indicative of learners who
have preferred those resources over and above others. For
III.I. HYPOTHESIS experimentation, we tune this threshold value, so that we get a
learner subset that is big enough to be tested for future steps.
In order to arrive at the formal algorithm design and answer
Step 2- Alg. "nominal_attributes_quantify"
the compound key question of our research, i.e., how to
input- number of learners using nominal attribute parameter
establish the link between learning resources and user profile
pair (p,q) in common
attributes that play a major role in learning(as discussed
output- numerical values for nominal attribute parameters
above), we formulate our research hypothesis as follows :
H1: Learning resources, apart from being distinguishable
 nominal attributes(a3, a4)
a. Construct matrix A5*5 with each row and
solely by learner ratings, can be more efficiently categorized
column representing specific nominal attribute
and tagged according to learner study motives.
parameters, and Aij = no. of users who have
H2: Incorporating user centric attributes for suggesting
parameter i and j in common.
customized learning pathways leads to pedagogically more
b. Apply NMF to split A into B & C:
A5*5=B5*k  Ck*5
accurate and satisfactory recommendations for the learner.
We address the first hypothesis in this paper and formalize it where Ck*5  features matrix (rows= features,
as follows: columns= nominal attribute parameters), and
Let ‘S’ be a System Cij= weight of each nominal attribute parameter
L={ l1, l 2, l 3, ....., lm } be a set of learning resources. to the feature extracted,
U={ u1, u 2, u 3, ...., un } be a set of learners. and B5*k  weight matrix (rows= nominal
A={ a1, a2, a3, a4, a5 } be a set of learner profile attributes. attribute parameters, columns=features), and
R={1,2,3,4,5,...,10} be set of possible ratings. Bij= weight of each feature in nominal attribute
where parameter
m, n  N c. Now,
a1  current skill level  {1,2,3,4,5,6} (i)  rows in B,
a 2  target skill level  {1,2,3,4,5,6},  ui U , a2>a1 A. Consider maximum(value).
(i=1 to n) B. Choose feature row in C, corresponding
a 3  learning strategy {1,2,3,4,5} to max (value).
/*nominal attribute*/ C. Obtain an ordering that defines how
a  preferred presentation style  {1,2,3,4,5}
4
similar (important) each nominal attribute
/*nominal attribute*/ parameter is to the other.
a  learning time N
5 /*discretized as different D. Construct matrix D where each row
ranges*/ identifies one linear ordering .
pij  user ui 's profile attribute aj (i=1 to n, j=1 to 5) (ii)Make matrix D symmetric by replacing Dij
with minimum(Dij, Dji).
riy  user ui 's rating of learning resource ly , (i=1 to
(iii)  rows in D, take average score
n, y=1 to m)
(Each average value represents possible
Problem Statement:
numerical values that the nominal attribute
 ly L, tag ly with each aj  A , (y=1 to m , j=1 to 5)
can take).
Fig (ii) Alg. "nominal_attributes_quantify"
Secondly, we apply NMF as shown in Fig (ii). NMF will Now, as shown in Fig (iii), in the subset of learners identified,
extract features from nominal attribute parameters and output we apply clustering to group learners based on user profile
how much importance does each feature have with respect to attributes(discussed in Sec II.I). Each learner is a five
the other parameters in terms of weights. And, because dimensional point and the coordinates represent his attributes.
nominal attribute parameters cannot be both similar to each We then pick the cluster having largest size. This cluster
other by different numerical metrics, we make matrix D further refines the users according to certain set of user profile
symmetric. The average of each linear ordering in matrix D attributes. Intuitively, it is a representative of certain pattern in
indicates an average similarity with other nominal attribute learning behaviour.
parameters. Thus, the output of this step is a numerical value
Step 4- Alg. "frequent_attributes_filter"
for each nominal attribute parameter, which can be used in the
input- data points of the largest cluster chosen
next step.
output- most frequently occurring user profile attributes
Step 3.1- Alg. "group_learner" Choose support level sl  {0,..,1}and sl  R
input- data points for every learner, with each coordinate /*association analysis*/
representing different profile attributes L1  {1 itemsets}
output- cluster with largest number of learners for(k=2,Lk-1   ,k++)
S { }
{
Pick first point pij at random
ck  {c|c=a  {b}  a  Lk-1  b  Lk-1  b  a}
S  S  { pij }
k
n=1
 ti  C /*for each learner record ti in the cluster*/
choose a value k N /*k=number of clusters*/
{
while n  k /*clusters chosen already*/
ct  {c|c  ck  c  ti}
{
Add point q ' to S, such that  c  ct, c.count++; /*maintain count of frequency*/
}
q '  pij and min{distance( q ' ,S)} > distance( qi ,S) Lk  {c  ck|c.count>sl} /*include learner itemsets with
( qi  pij - q ' , i=1 to n, j=1 to 5) count greater than min. support level*/
S S  { q'} }
n=n+1 return L= Lk
} k

/*returns most frequently occurring learner profile attributes*/


Step 3.2
While any point has changed in cluster made: Fig (iv) Alg. "frequent_attributes_filter"
/*K nearest neighbour based clustering*/
{ Next, the focus is to quantify this pattern and address the
 pij (i=1 to n,j=1 to 5) question of tagging learning resources based on the largest
 Sa  S (a=1 to k) cluster chosen. The solution to this problem is application of
d(Sa, pij )=  Sa1  Pi1 2 ,.......,   Sa 5  Pi 5 2 association analysis to extract hidden knowledge and
interesting relationships from this processed data(Fig(iv)). We
/*Euclidean distance metric*/ use frequent itemsets to find the most commonly attributes that
x[a]=min{d(Sa, pij )} occur together using the apriori principle. A minimum support
get l=a which corresponds to min(x[]) level is defined as an input to this fourth step. The output is a
Sl=Sl  pij collection of profile attributes, among which we pick the one
 every Sl ,calculate mean of points (M) in that cluster with largest frequency. Finally, we assign this most frequently
Sl  Mean(M) occurring and largest itemset of attributes to the learning
} resource.
Step 3.3
Reduce ‘k’ till average diameter D increases by a very large
factor /*choose optimal no. of clusters*/
D= max(d(p, q))  (p, q)  pij V. EXPERIMENTAL RESULTS AND DISCUSSION
Step 3.4 The main contribution of our work is the application of
Choose cluster with largest number of attribute itemsets machine learning techniques to develop a novel system that
C  max{Size(Sa )} can aid in more accurate learning resource pathways
recommendation. The strength of the system lies in assisting
Fig (iii) Alg. "group_learner"
new learners by providing more relevant and structured
navigational support for informal education settings. Instead of
"People similar to you who liked learning resource X also
liked Y", recommendations take the form of "People similar Some may wonder that it's difficult to evaluate such a
to you with learning strategy 3, current skill level 2, target skill system. But, the identified way of evaluating this randomized
level 5, preferred presentation style 4, available learning time simulation is by implicit inference of whether the target learner
[20-30] time units who liked the learning resource X also liked was recommended learning resources based on full user profile
Y". The advantages are: cloud of similar learners and in the same domain. The only
shortcoming is that we cannot explicitly infer user satisfaction
1)Weight of recommendation and it's specificity is increased.
and correctness of the proposed learning pathway through a
2)A sense of faith in the recommender system is introduced
questionnaire. However, the importance and relevance of
because peer learners having similar competencies are
implicit inference lies in the fact that it can be collected more
matched.
easily and at lower cost to the user, although inferences about
3)The learner is placed in a more comfortable and confident
learning resource desirability will be slightly less accurate than
position to choose his learning resources.
explicitly supplied feedback [15].
4)Also, learning resources get organized according to learner
characteristics. Next, we grouped the subset of users, who rated learning
For the implementation, we used the "Book Crossing" resources greater than 6(on a scale of 10). Using intelligent
dataset [14], consisting of 2,78,858 learners providing multi attribute visualization techniques, we plotted the
1,149,780 ratings about 2,71,379 learning resources(books). clustering results for learning resource with ISBN
Because TEL datasets having additional and explicit "000649840X" and ISBN "0684867621" in the dataset, and
information about learning attributes are very rare to find, the came up with interesting observations. In Fig (vii), each
implementation phase consisted of generating random values attribute is represented in a vertical line, where the maximum
for various user profile attributes, as mentioned in Section and minimum values of that dimension are scaled to the upper
II.II, and augmenting them with the "Book Crossing" dataset. and lower points on these vertical lines. For 5 visualized
Then, on applying NMF and extracting 10 independent attributes, 4 lines connected to each vertical line at the
features, we achieved the following results for the two nominal appropriate dimensional value represent a 5-dimensional point.
attributes: The plot shows few prominent horizontal lines(data points),
which intuitively represent strong candidates for tagging
learning resources. Coming to Fig (viii), it provides further
enhancements to the visualization by depicting data points that
are attracted to anchors with value dependent positions. The
data instances are placed inside a 5-sided polygon because the
clustering process is based on 5 different learner profile
attributes.

Fig (v)Learning strategy

Fig (vi)Presentation style


Fig (vii)Parallel Coordinates Visualization
From Fig (v) and Fig (vi), we infer that learning strategy 3 and ( for ISBN "000649840X")
5 are most similar or important with respect to each other,
while learning strategy 4 and 5 are least similar. Also, among
the preferred presentation styles, 4 and 5 are most similar(least
distance between them), while 1 and 3 are the least
similar(maximum distance).
V. RELATED WORK
In this section, we describe the prior studies that have been
made in the context of development of learner models for
educational adaptive hypermedia. In [17], Sicilia et al. carried
out an empirical study on the Merlot repository dataset and
concluded that ratings cannot correctly differentiate learner
preferences. Results showed relatively high mean absolute
errors of around 1.0 in a scale of 1 to 5. Psychological studies
have stressed on the need to include personalization factors
like collective intelligence, competencies of peer learners and
learner's reaction towards the learning objects recommended,
for inclusion in the design of TEL RecSys [18]. Instead of
focusing on development of better algorithms, such learner
centric factors can improve accuracy, as learners are both
producers of data and recipients of information. A theoretical
Fig (viii)Polyviz Visualization(for ISBN "0684867621") study on learner models like stereotype, overlay, differential,
perturbation and plan models was made in [21], and a generic
Next, on applying association analysis to the data points of approach to build these models was provided. However, no
the largest cluster, we get the certain tags based on most algorithms or implementation details were described.
frequently occurring attribute itemsets. Keeping the support In [19], Verbert et al. did a survey on context aware
level constant at 0.1, we generate tags for each of the learning recommender systems and classified the context framework of
resources, four of which are shown in Fig (ix). If there are TEL RecSys on the basis of computing, location, time,
many tags having same frequencies, we generate tag clouds for activity, physical conditions, resource and users. The study
the corresponding learning resources, where each cloud can showed that less than half of the TEL RecSys actually took
hold more than one kind of preference(tag). computing context into consideration. For a majority of them,
Though some of the current tags do not include all 5 user the implementation was still in the prototype phase. Learner
profile attributes, we claim that the system will generate better characteristics like basic personal information, prior
and higher order tags, once more number of learners start knowledge level(performance), interests, learning goals,
using the system. This is because the input learner subset will cognitive style and affects(emotions) were identified as key
grow in size at each step of the system and this will increase points for consideration in providing personalized learning
the number of possible combinations of learner profile services. In [12], these determinants were implemented in
attributes. The justification seems valid, because our proposed NetLogo simulation environment, with an objective of
system is a bottom up system which incrementally improves as identifying the preference and competence gap of learners.
learners and their ratings increase. It is feasible for lifelong Formulas for implicit and explicit learning variables laid the
learning scenarios with changing learning actions. basis of determining the percentage of graduates, their
satisfaction levels and time taken to graduate. However, to the
Learning Resource Tag-Cloud generated best of our knowledge, the identified user profile attributes
ISBN have not been used for improving pedagogical accuracy of
000649840X [6, 6, [41-50], 24240, 20549] learning pathways. And because this a relatively new area, not
much work has been done in applying machine learning
002542730X [4, 5, -, 30380,23783] oriented approaches to leverage such user tagged parameters.
and So, it is not trivial to use the previous work along this line to
[-, 6, [51-60], 24240, 22802] directly serve as subcomponents in our setting.
038550120X [6, 6, [51-60], 24240, 23783]
and
[6, 6, [21-30], 17608, 23783] VI. CONCLUSION
0060928336 [-, -, -, 30380,22802]
and In this paper, we have designed a solution to cater to learning
[-, 6, [41-50], -, -] resource recommendations that inaccurate and less relevant.
This disturbs learners and wastes their precious time in
Fig(ix)Tags generated for learning resources(5 fields in tag searching for the apt learning resources. We call this area of
cloud are Current Skill level, Target skill level, Learning technology enhanced learning sensitive because today, many
time(range in hrs), Learning strategy & Presentation style) people from different demographics having different
competencies, different ways of learning and interpretation,
collaborate on online education portals and leverage resources
from learning repositories to improve their skill set. Because [7] Based on: Merlot Pedagogy portal( http://pedagogy.merlot.org/)
[8] The CDIO Syllabus: A comparative study of expected student
of the inefficiency of human support that can be provided to
proficiency", European Journal of Engineering Education, Vol.28 No. 3
such a large learning network, intelligent ways are required to (2003)
match learner's actual preferences with learning resources. And [9] Based on: Studies of University of Kansas Center for Research on
leveraging user specific learning attributes is one of the ways Learning(http://www.kucrl.org/sim/strategies.shtml)
to model real life learner behaviour and improve the [10] D. Dagger, O. Conlan, and V. Wade. Fundamental requirements
of personalised e learning development environments. In G. Richards,
pedagogical accuracy of suggested learning pathways. editor, Proceedings of World Conference on E-Learning in Corporate,
Government, Healthcare, and Higher Education 2005, pages 2746–
2754. AACE, October 2005
[11] WALKER A, Recker M M, Lawless K and Wiley D (2004).
VII. FUTURE WORK Collaborative Information Filtering: a review and an educational
application. International Journal of Artificial Intelligence in
For future work, we aim to address hypothesis H2, as Education, 14, 1-26.
explained in Section III.I. We also plan to evaluate our system [12] Nadolski, Rob J., Bert Van den Berg, Adriana J. Berlanga, Hendrik
on "SNAP Amazon product co purchasing network" metadata Drachsler, Hans GK Hummel, Rob Koper, and Peter B. Sloep.
"Simulating light-weight personalised recommender systems in learning
[13], because it is more informative than the current dataset networks: A case for pedagogy-oriented and rating-based hybrid
used in experimentation. Furthermore, as the system becomes recommendation strategies." Journal of Artificial Societies and Social
bigger, accuracy could be tested by dynamically increasing Simulation 12, no. 1 (2009): 4.
support levels and getting top 'k' patterns of the highest order. [13] J. Leskovec, L. Adamic and B. Adamic. The Dynamics of Viral
Marketing. ACM Transactions on the Web (ACM TWEB), 1(1), 2007.
The tags achieved at higher support levels would be better than [14] Improving Recommendation Lists Through Topic Diversification, Cai-
the ones at lower support levels. The goal is to choose an Nicolas Ziegler, Sean M. McNee, Joseph A. Konstan, Georg
optimal support level, so that order of attribute itemsets is as Lausen; Proceedings of the 14th International World Wide Web
high as possible, and the number of those itemsets having same Conference (WWW '05), May 10-14, 2005, Chiba, Japan.
[15] Resnick, P. & Varian, H. (1997). Recommender systems, Special Issue.
frequency is as small as possible. This will reduce ambiguity Communications of the ACM, 40(3), 56-58
and control the trade off between which and how many [16] Manouselis, Nikos, Riina Vuorikari, and Frans Van Assche. "Simulated
attribute itemsets to use as a tag for the learning resource. analysis of MAUT collaborative filtering for learning object
Another possible research direction is to extend this simulation recommendation." In Proceedings of the 1st Workshop on Social
Information Retrieval for Technology Enhanced Learning. 2007.
based model to develop real world datasets for TEL RecSys. [17] Sicilia, Miguel-Ángel, Elena García-Barriocanal, Salvador Sánchez-
Such an augmented labelling of learning resources will be Alonso, and Cristian Cechinel. "Exploring user-based recommender
more reflective of learner proficiencies and objectives. results in large learning object repositories: the case of MERLOT."
Procedia Computer Science 1, no. 2 (2010): 2859-2864.
[18] Buder, Jürgen, and Christina Schwind. "Learning with personalized
ACKNOWLEDGMENT recommender systems: A psychological view." Computers in Human
Behavior 28, no. 1 (2012): 207-216.
[19] Verbert, Katrien, Nikos Manouselis, Xavier Ochoa, Martin Wolpers,
This work was supported by the National Research Foundation Hendrik Drachsler, Ivana Bosnic, and Erik Duval. "Context-aware
of Korea(NRF) grant funded by the Korea government(MEST) recommender systems for learning: a survey and future challenges."
(2012): 1-1.
(No. NRF-2013R1A1A2013401).
[20] Brusilovsky, Peter, and Eva Millán. "User models for adaptive
hypermedia and adaptive educational systems." In The adaptive web, pp.
3-53. Springer-Verlag, 2007
REFERENCES [21] Nguyen, Loc, and Phung Do. "Learner model in adaptive
[1] Colley, H., Hodkinson, P., & Malcolm, J. (2002b). Non-formal learning: learning." World Academy of Science, Engineering and Technology 45
mapping the conceptual terrain. a consultation report [Electronic (2008): 395-400.
Version]. Retrieved 01.05.2008 from http://www.infed.org/archives/e- [22] Hartigan, John A., and Manchek A. Wong. "Algorithm AS 136: A k-
texts/colley_informal_learning.htm means clustering algorithm." Journal of the Royal Statistical Society.
[2] Manouselis, Nikos, Hendrik Drachsler, Riina Vuorikari, Hans Hummel, Series C (Applied Statistics) 28, no. 1 (1979): 100-108.
and Rob Koper. "Recommender systems in technology enhanced [23] R. Agrawal and R. Srikant, “Fast Algorithms for Mining Association
learning." In Recommender systems handbook, pp. 387-415. Springer Rules,” Proceedings of the 20th International Conference on Very Large
US, 2011. Data Bases,1994, pp. 487-499.
[3] Mccalla G(2004), The ecological approach to the design of e-Learning [24] Lee, Daniel D., and H. Sebastian Seung. "Learning the parts of objects
environments:Purpose-based capture and use of information about by non-negative matrix factorization." Nature 401, no. 6755 (1999): 788-
learners. Journal of Interactive Media in Education. Retrieved October 791.
18, 2007, from http://www-jime.open.ac.uk/2004/7 . [25] A. Riddell. Factors influencing educational quality and effectiveness in
[4] Watson, William R., Sunnie Lee, and Charles M. Reigeluth. "Learning developing countries: A review of research. Deutsche Gesellschaft fur
Management Systems: An Overview and Roadmap of the Systematic Technische Zusammenarbeit (GTZ), Germany, 2008
Application of Computers in Education." Advances in computer- [26] Knowledge for Development: World Development Report 1998/99.
supported learning (2007): 66-96 World Bank, 1998.
[5] Huang, Y.-M., Huang, T.-C., Wang, K.-T., Hwang, W.-Y. (2009). A
Markov-based Recommendation Model for Exploring the Transfer of
Learning on the Web. Educational Technology & Society, 12(2), 144-
162.
[6] Borich, G. D., & Tombari, M. L. (1995). Educational Psychology: A
Contemporary Approach, New York: Harper Collins College.

You might also like