Professional Documents
Culture Documents
Habibi TMP Assetyx5x6z
Habibi TMP Assetyx5x6z
Habibi TMP Assetyx5x6z
wCONICET, Consejo Nacional de Investigaciones Cientı́fı́cas y Técnicas, Av. Rivadavia 1917, Capital Federal (CP1033), Bs. As., Argentina
Abstract People have unique ways of learning, which may greatly affect the learning process and,
therefore, its outcome. In order to be effective, e-learning systems should be capable of
adapting the content of courses to the individual characteristics of students. In this regard,
some educational systems have proposed the use of questionnaires for determining a student
learning style; and then adapting their behaviour according to the students’ styles. However,
the use of questionnaires is shown to be not only a time-consuming investment but also an
unreliable method for acquiring learning style characterisations. In this paper, we present an
approach to recognize automatically the learning styles of individual students according to
the actions that he or she has performed in an e-learning environment. This recognition
technique is based upon feed-forward neural networks.
Introduction presented in the way that best fit the learning style of
each student, which is usually assessed through a
Students learn in different ways. Some students prefer
predefined questionnaire. However, answering long
graphics, such as diagrams and blueprints, while oth-
questionnaires is a time-consuming task that students
ers prefer written material. Some students feel more
are not always willing to carry out and, consequently,
comfortable with facts, data and experimentation,
results become unreliable (Stash et al. 2004).
whereas others prefer principles and theories (Felder
This paper describes an approach to the problem of
& Silvermann 1988). E-learning environments can
mapping students&apso; actions within e-learning
take advantage of these different forms of learning by
environments into learning styles. The method is
recognizing the style of each individual student using
based on artificial neural networks (ANNs). Neural
the system and adapting the content of courses to
networks are computational models for classification
match this style.
inspired by the neural structure of the brain: models
There are a few systems that are actually capable of
that have proven to produce very accurate classifiers.
adapting courses’ contents according to students’
In the proposed approach, feed-forward neural net-
learning styles (Carver et al. 1999; Gilbert & Han
works are used to recognize students’ learning styles
1999; Paredes & Rodriguez 2002; Stash & Brau
based upon the actions they have performed in an
2004). In these systems, the learning materials are then
e-learning system.
The rest of this paper is organized as follows: The
Accepted: 16 March 2006
next section briefly describes learning style theory.
Correspondence: Jorge E. Villaverde, ISISTAN Research Institute,
UNICEN University, Campus Universitario, Paraje Arroyo Seco, Tandil This is followed by an overview of ANNs and the
(CP 7000), Bs. As., Argentina. E-mail: jvillave@exa.unicen.edu.ar back-propagation algorithms. The sections after that
& 2006 The Authors. Journal compilation & 2006 Blackwell Publishing Ltd Journal of Computer Assisted Learning 22, pp197–206 197
198 J. E. Villaverde et al.
show how we have modelled the recognition of practical, and are oriented towards facts and
learning styles using neural networks and empirical procedural information. When solving problems,
results obtained in the recognition of learning styles. sensory students are routinely very patient with
The penultimate section discusses related work and details and usually dislike surprises. Because of
the last section presents our conclusions. these characteristics, they show a slower reaction
to problems, but they typically present a better
outcome.
Learning styles Intuitive learners: intuitive learners prefer the-
Keefe (1979) defines learning styles as characteristic ories and principles. They rapidly become bored
cognitive, affective and psychological behaviours that with details and mechanical problem solving.
serve as relatively stable indicators of how learners Innovation is what attracts intuitive learners’ at-
perceive, interact with and respond to learning en- tention. They generally solve problems quickly,
vironments. In the last decade, several learning styles not paying much attention to details. This makes
models have been proposed (Kolb 1984; Myers & them fast but prone to errors and, then, they
McCaulley 1985; Felder & Silvermann 1988; Lawr- often get lower qualifications than sensitive
ence 1993; Litzinger & Osif 1993). A critical review learners.
of learning styles, analysing their reliability, validity
and implication for pedagogy, can be found in (Cof- Input: Through which sensory channel is external
field et al. 2004a;b). The authors of this review con- information most effectively perceived: visual——
cluded that in the field of learning styles, there is a lack pictures, diagrams, graphs, demonstrations or ver-
of theoretical coherence and a common framework bal——written or spoken sounds?
(Coffield et al. 2004b, p. 145).
Visual learners: they remember, understand and
In spite of this fact, experimental research on the
assimilate information better if it is presented to
application of learning styles in computer-based edu-
them in a visual way. They tend to remember
cation provides support for the view that learning can
graphics, pictures, diagrams, time lines, blue-
be enhanced through the presentation of materials that
prints, presentations and any other visual material.
are consistent with a student’s particular learning style
Verbal learners: cognitive scientists have estab-
(Budhu 2002; Peña et al. 2002; Stash et al. 2004). For
lished that our brains generally convert written
example, it has been shown that the performance of
works into their spoken equivalents and process
students in a simple Web-based learning environment
them in the same way that they process spoken
correlates with their self-reported learning preference
words (Felder & Brent 2005). Hence, verbal
(Walters et al. 2000).
learners are not only those who prefer auditory
In this paper, we have adopted the model suggested
material but also those who remember well what
by Felder and Silverman (1988) for engineering edu-
they hear and what they read.
cation, which classifies students according to their
position in several scales that evaluate how they per- Processing: How does the student prefer to process
ceive and process information. information: actively——through engagement in
This model classifies students according to four physical activities or discussions, or reflectively——
dimensions: through introspection?
Perception: What type of information does the Active learners: they feel more comfortable with
student preferentially perceive: sensory (ex- active experimentation than with reflexive ob-
ternal)——sights, sounds, physical sensations, or servation. An active person learns by trying
intuitive (internal)——possibilities, insights, hun- things out and working with others. They like
ches? doing something in the external world with the
received information. Active learners work well
Sensory learners: these students like facts, data in groups and in situations that require their
and experimentation. They perceive concrete, participation.
& 2006 The Authors. Journal compilation & 2006 Blackwell Publishing Ltd
Learning styles’ recognition in e-learning 199
& 2006 The Authors. Journal compilation & 2006 Blackwell Publishing Ltd
200 J. E. Villaverde et al.
Structure and operation of ANNs of the most spread models in the connectionism field,
due to its learning capability to associate an input
The fundamental unit of processing in ANNs is the
space to an output. It has been demonstrated by dif-
Artificial Neuron (AN); see Fig 1. ANs are mathe-
ferent authors that a multi-layer perceptron is a uni-
matical models that represent the general character-
versal function approximation mechanism, in the
istics of biological neurons. Let us see how these
sense that any continuous function over a compact Rn
biological features are mathematically modeled and
can be approximated by a multi-layer perceptron
how they contribute to the learning process of ANNs.
(Cybenko 1989; Hornik et al. 1989).
When a signal is transmitted from processing unit i
BPN nets operate in two steps. In the first step,
to processing unit j, the signal xi is modified by the
called training process, the network is initialized with
synaptic weight (wij ) associated with this communica-
random small values in its weights. The goal of this
tion channel. The modulated signals that arrive at unit
process is to find a set of weight values that minimize
j are added to form the net input Netj as is shown in
X the global error of the network, given in equation (5).
Netj ¼ xi wij ; ð1Þ The second step is called generalization. In this step,
i
the network has already learned an internal re-
Each neuron is characterized in any instant of time t presentation of the previously presented patterns and
by a real value, called activation state or activation becomes able to classify novel patterns presented as
level, aj ðtÞ Also, there is a function F, called activa- inputs.
tion function, which determines the next activation The learning process of a BPN is briefly described
state based on the current activation state and the Netj as follows: a pattern is a two-tuple Pi ¼ hXi ; Ti i where
input of the neuron. Xi is a set of values that will be presented as input to
F aj ðtÞ; Netj ¼ aj ðt þ 1Þ: ð2Þ the network and Ti is a set of values that represent the
desired target for the values presented at the input
Associated with each unit there exists an output
layer. For the network to learn a pattern, Pi, the values
function, fj , that transforms the current activation level
of the pattern have to be presented at the input layer
into an output signal yi. This signal is sent through a
first. These values are taken as stimulus and are pro-
unidirectional communication channel to other units in
pagated forward until the output layer is reached. In
the network
the output, a set (O) of values obtained by the network
fj aj ðt þ 1Þ ¼ yj ð3Þ (called oj) are compared with the set of desired values
(Ti) to obtain the pattern error according to
Learning algorithm X
M 2
Erri ¼ oj tj ; ð4Þ
A learning algorithm is the process by which an ANN j¼1
& 2006 The Authors. Journal compilation & 2006 Blackwell Publishing Ltd
Learning styles’ recognition in e-learning 201
1986;1988) and is a generalization of the Widrow– per observed action in the system. These actions are as
Hoff delta rule (Widrow & Hoff 1960) follows:
& 2006 The Authors. Journal compilation & 2006 Blackwell Publishing Ltd
202 J. E. Villaverde et al.
Table 1. Input representation. Although there are some empirical rules, such as the
Baum–Haussler rule (Baum & Haussler 1988), for
X Action 5 15
determining the desirable number of neurons in the
x0 Reading material Abstract Concrete hidden layer of a multi-layer feed-forward neural
x1 Access to examples Few Much network, we determined this architectural parameter
x2 Answer changes Few Much via trial-and-error experimentation. A total of 24 units
x3 Exercises Few Much
have been empirically found as appropriate for the
x4 Exam delivery time Quick Slow
x5 Exam revision Few Much
task at hand, considering that this layer has to have
x6 Chat usage Don’t use Read and write enough processing units to represent the nonlinear
x7 Forum usage Don’t use Read and post aspects of the model and not too many for making the
x8 Mail usage Don’t use Use training process very complex.
x9 Information access Lineal Global
& 2006 The Authors. Journal compilation & 2006 Blackwell Publishing Ltd
Learning styles’ recognition in e-learning 203
Table 3. Proposed Architectural Parameters. relevance of past experience in the system to the re-
cognition of the learning style.
Parameter Value
Experimental results
Number of input neurons 10
Number of hidden neurons 24 Estimating the accuracy of classifiers induced by su-
Number of output neurons 3 pervised learning algorithms (i.e. classifier’s prob-
Activation function Hyperbolic tangent
ability of correctly recognizing a randomly selected
Learning rate 0.02
Momentum 0.5
instance) is important not only to predict its future
prediction accuracy but also for choosing a classifier
from a given set (model selection) or combining
To sum up, the proposed architecture for learning classifiers (Wolpert 1992). In order to evaluate the
style recognition is a three-layered feed-forward proposed approach, we generated an artificial data-
neural network with Batch Gradient Descent with the set for experimentation by simulating the actions of
Momentum learning algorithm. In this architecture, students.
the first layer contains a total of 10 input neurons; the For this task, we considered that each student has a
hidden layer contains 24 processing units that are particular learning style denoted by a set of preferred
connected to 3 output units in the third layer. In- actions and behaves according to it. The resulting
formation about the network architecture and other dataset built for testing the network consisted of
parameters is summarized in Table 3. 100 pairs of input–output values. Each of these pairs
contained possible actions that students might perform
Temporal adaptation in the system associated with their corresponding
learning style dimensions.
To ensure that the student’s learning style does not
To determine the best number of processing units in
oscillate between different states (and keeps the sys-
the hidden layer, we trained the network varying this
tem changing the contents of the student’s courses as a
parameter and estimated the architecture accuracy
consequence), historical interaction data can be sup-
using k-fold cross-validation with k 5 10. Each of the
plied to the classifier to enhance the prediction of the
10 folds was composed of a set of 90 training patterns
learning style. To accomplish this goal, a parameter N
and 10 test patterns. To measure the accuracy of each
is defined to represent the length of the input sequence
classifier, we calculated the global error (equation (5))
to the network, causing the input to the classifier to be
that it produces when it is stimulated with the test
a function of the historical data. That is, the input to
cases. Each of the output dimensions was considered
the system is a finite sequence X where
independently in the calculation.
X ¼ xt ; xt1 ; . . . ; xtN : ð11Þ Figure 2 shows the mean accuracy of cross-valida-
tion when varying the number of neurons in the net-
work-hidden layer. We can observe in the figure that
This sequence of inputs is pre-processed applying
the best accuracy (69.3%) is reached when the number
the discount return transformation (Sutton 1988) be-
of hidden units is 24. Figure 3 shows the variance of
fore becoming the real input to the network. Equation
the classifier when the number of neuron varies. In this
(12) shows how the actual input to the ANN will be
figure, the lowest variance (1.53%) is obtained when
produced from the data measured at time t and from
the number of neurons is 22. Based on these results, we
historical information
have chosen 24 neurons for achieving the maximum
X
N
accuracy. For this value, the variance (1.70%) of the
xt ¼ gk1 xtþk 0 g 1: ð12Þ
classifier is not the minimum but it is still acceptable.
k¼1
& 2006 The Authors. Journal compilation & 2006 Blackwell Publishing Ltd
204 J. E. Villaverde et al.
68.6
AHA! (Stash & Brau 2004), an educational system
that infers the students’ learning style from the ob-
68.4
servation of their browsing behaviour. In contrast to
68.2
the mentioned systems, it does not make use of
68 questionnaires. However, this system covers a small
67.8 number of learning styles, those corresponding to
67.6 field-dependent/independent styles and to visual/ver-
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
# Neurons
bal styles, whereas the approach presented in this pa-
per is based on the Felder and Silverman theory for
Fig 2 Mean accuracy as a function of the number of neurons in engineering education.
the hidden layer of the network. Learning styles in engineering education have been
studied by (Al-Holou & Clum 1999; Ayre & Nafalski
1.9 2000; Budhu 2002; Kuri & Truzzi 2002). Kuri and
1.85 Truzzi (2002) use the Felder–Silverman model to as-
sess the distribution of responses of freshmen en-
1.8
gineering students using the ILS questionnaire, and
1.75 they concluded that the use of learning styles to ac-
Variance
& 2006 The Authors. Journal compilation & 2006 Blackwell Publishing Ltd
Learning styles’ recognition in e-learning 205
system usage so that systems using this approach can Gilbert J.E. & Han C.Y. (1999) Arthur: adapting instruction
recognize changes in learning styles or some of their to accommodate learning style. In World Conference on
dimensions over time. the WWW and Internet (WebNet 99) (eds P.D. Bra & J.J.
The recognition mechanism presented in this paper Leggett), pp. 433–438. Honolulu, Hawaii, USA.
can be introduced in adaptive e-learning environments Hornik K., Stinchcombe M. & White H. (1989) Multilayer
to help in the detection of students’ learning styles feedforward networks are universal approximators.
Neural Networks 2, 359–366.
and, thus, conveniently adapt the contents of academic
Keefe L.W. (1979) Learning style: an overview. In NASSP’s
courses that are presented to them. It can also be ex-
Student learning styles: Diagnosing and prescribing
tended to consider further input actions available in
programs, pp. 1–17.
particular e-learning systems or domains.
Kim K.S. & Moore J.L. (2005) Web-based learning: factors
affecting students’ satisfaction and learning experience.
References First Monday 10. Available from http://www.firstmonday.
dk/issue/issue10-11/kim/index.html
Al-Holou N. & Clum J.A. (1999) Teaching electroscience Kolb D.A. (1984) Experiential Learning: Experience as the
using computer-based instruction. In 29th ASEE/IEEE Source of Learning and Development. Prentice-Hall,
Frontiers in Education Conference (eds D. Budny &
New York.
G. Bjedov), pp. 24–29. San Juan, Puerto Rico.
Kuri N.P. & Truzzi O.M.S. (2002) Learning styles of
Ayre M. & Nafalski A. (2000) Recognising diverse learning
freshmen engineering students. In International Con-
styles in teaching and assessment of electronic en-
ference on Engineering Education (ICEE 2002), Man-
gineering. In 30th ASEE/IEEE Frontiers in Education
chester, UK.
Conference (eds D. Budny & G. Bjedov), pp. 18–23.
Lawrence G. (1993) People Types and Tiger Stripes: A
Kansas City, MO.
Practical Guide to Learning Styles. Center for Applica-
Baum E.B. & Haussler D. (1988) What size net gives valid
tions of Psychological Type. Gainesville, FL.
generalization? Neural Computation 1, 151–160.
Litzinger M.E. & Osif B. (1993) Accommodating diverse
Bishop C.M. (1995) Neural Networks for Pattern Recogni-
learning styles: designing instruction for electronic infor-
tion. Clarendon Press, Oxford.
mation sources. In What is Good Instruction Now? Libr-
Budhu M. (2002) Interactive web-based learning using in-
ary Instruction for the 90s (ed. L. Shirato). Pierian Press.
teractive multimedia simulations. In International Con-
McCulloch W.S. & Pitts W. (1943) A logical calculus of
ference on Engineering Education (ICEE 2002), pp. 1–6.
ideas immanent in nervous activity. Bulletin of Mathe-
International Network for Engineering Education & Re-
search, Manchester, UK. matical Biophysics 5, 115–133.
Carver C.A.J., Howard R.A. & Lane W.D. (1999) Enhancing Myers I. & McCaulley M. (1985) Manual: A Guide to the
student learning through hypermedia courseware and Development and Use of the Myers–Briggs Type In-
incorporation of student learning styles. IEEE Transac- dicator. Consulting Psychologists Press., Palo Alto, CA.
tions on Education 42, 33–38. Paredes P. & Rodriguez P. (2002) Considering learning
Coffield F., Moseley D., Hall E. & Ecclestone K. (2004a) styles in adaptive web-based system. In Proceeding of
Learning styles and pedagogy in post-16 learning: a the 6th World Multiconference on Systemics, Cybernetics
systematic and critical review. Technical Report 041543, and informatics (eds N. Callaos, M. Margenstern &
Learning and Skills Research Centre. B. Sanchez), pp. 481–485. Orlardo, FL.
Coffield F., Moseley D., Hall E. & Ecclestone K. (2004b) Parker D.B. (1982) Learning logic, invention report. Tech-
Should we be using learning styles? What research has to nical Report S81–64, File 1, Office of Technology Li-
say to practice. Technical Report 041540, Learning and cencing, Stanford University.
Skills Research Centre. Peña C.I., Marzo J.L. & de la Rosa J.L (2002) Intelligent
Cybenko G. (1989) Approximations by superposition of agents in a teaching and learning environment on the
sigmoidal functions. Mathematics of Control, Signals, Web. In ICALT2002. IEEE, Kazan, Russia.
and Systems 2, 303–314. Rosenblatt F. (1958) The perceptron: a probabilistic model
Felder R.M. & Brent R. (2005) Understanding student dif- for information storage and organization in the brain.
ferences. Journal of Engineering Education 94, 57–72. Psychological Review 65, 386–408.
Felder R.M. & Silvermann L.K. (1988) Learning and Rumelhart D.E., Hinton G.E. & Williams R.J. (1986)
teaching styles in engineering education. Journal of En- Learning representations by back-propagating errors.
gineering Education 78, 674–681. Nature 323, 533–536.
& 2006 The Authors. Journal compilation & 2006 Blackwell Publishing Ltd
206 J. E. Villaverde et al.
Rumelhart D., Hinton G. & Williams R. (1988) Learning Sutton R.S. (1988) Learning to predict by the methods of
internal representations by error propagation. In Parallel temporal differences. Machine Learning 3, 9–44.
Distributed Processing: Explorations in the Micro- Walters D., Egert C. & Cuddihy E. (2000) Learning styles and
structure of Cognition (eds D. Rumelhart & G. McClel- web-based education: a quantitative approach. In Pro-
land), pp. 318–362. MIT Press, Cambridge, MA. ceedings from the 9th Annual FACT Conference on In-
Stash N. & Brau P.D. (2004) Incorporating cognitive styles structional Technology, pp. 115–117. Malo, Buffalo, NY.
in AHA! (The adaptive hypermedia architecture). In Werbos P.J. (1988) Generalization of backpropagation with
IASTED International Conference (ed. V. Uskov), pp. application to a recurrent gas market model. Neural
378–383. ACTA Press, Innsbruck, Austria. Networks 1, 339–356.
Stash N., Cristea A. & Brau P.D. (2004) Authoring of learn- Widrow B. & Hoff M.E. (1960) Adaptive switching circuits.
ing styles in adaptive hypermedia: problems and In IRE WESCON Convention Record, pp. 96–104, New
solutions. In 13th International Conference on York, NY.
World Wide Web – Alternate Track Papers & Posters, Wolpert D.H. (1992) Stacked generalization. Neural Net-
pp. 114–123. works 5, 241–259.
& 2006 The Authors. Journal compilation & 2006 Blackwell Publishing Ltd