ScienceDirect

ScienceDirect
Procedia Computer Science 00 (2018) 000–000
Procedia Computer Science 00 (2018) 000–000 www.elsevier.com/locate/procedia
Available online at www.sciencedirect.com
www.elsevier.com/locate/procedia

ScienceDirect
Procedia Computer Science 132 (2018) 834–842

International Conference on Computational Intelligence and Data Science (ICCIDS 2018)

International Conference on Computational Intelligence and Data Science (ICCIDS 2018)
Smart Education with artificial intelligence based determination of
Smart Education with artificial intelligence
learning styles based determination of
learning styles
Richa Bajajaa*, Vidushi Sharmabb
a,b
Richa Bajaj *, Vidushi Sharma
Department of ICT, Gautam Buddha University, Greater Noida 201302, India
a,b
Department of ICT, Gautam Buddha University, Greater Noida 201302, India

Abstract
Abstract
The need of the hour in present day education environment is adaptivity. Adaptive educational systems aim to customize content
and
The learning paths
need of the of in
hour students.
present These aid’s in environment
day education the minimizing disorientation
is adaptivity. and cognitive
Adaptive educationaloverload
systems problems; thus maximizing
aim to customize content
learning efficiency.
and learning paths ofPresent learning
students. Thesesystems
aid’s inarethelacking adaptivity;
minimizing as they offer
disorientation andsame resources
cognitive for all
overload users irrespective
problems; of their
thus maximizing
individual
learning needs and
efficiency. preferences.
Present learningStudents
systems learn according
are lacking to theiraslearning
adaptivity; they offerstyles
sameand determining
resources for all these is a crucial of
users irrespective step in
their
making eLearning
individual needs andor preferences.
traditional education adaptive.
Students learn To determine
according styles, and
to their learning styles learning models these
determining have isbeen suggested
a crucial step in
literature, but there or
making eLearning is no readily available
traditional educationsoftware
adaptive.toolTothat provideslearning
determine the flexibility
styles, to select and
learning modelsimplement the most
have been suitable
suggested in
learning model.
literature, To fulfil
but there is no this dire available
readily need, a framework
software toolof a that
tool provides
is proposedthehere, whichtotakes
flexibility into
select andconsideration
implement themultiple
most learning
suitable
models and
learning artificial
model. intelligence
To fulfil techniques
this dire need, for determining
a framework of a tool is students’
proposedlearning styles.takes
here, which Theinto
toolconsideration
would provide the facility
multiple learningto
compareand
models learning models,
artificial to determine
intelligence the most
techniques suitable onestudents’
for determining for a particular
learning environment.
styles. The tool It iswould
suggested
providethatthethis tool be
facility to
deployed in a cloudmodels,
compare learning environment to provide
to determine thea scalable solution
most suitable onethat
foroffers easy andenvironment.
a particular rapid determination of learning
It is suggested styles.
that this tool be
deployed in a cloud environment to provide a scalable solution that offers easy and rapid determination of learning styles.
© 2018
© 2018 The
The Authors. Published by
Authors. Published by Elsevier
Elsevier B.V.
Ltd.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/3.0/)
Peer-review
© 2018 The Authors. Published by Elsevier B.V. committee of the International Conference on Computational Intelligence and
under responsibility of the scientific
Peer-review under responsibility of the scientific committee of the International Conference on Computational Intelligence and
Data
Data Science
Peer-review (ICCIDS
Scienceunder 2018).
responsibility
(ICCIDS 2018). of the scientific committee of the International Conference on Computational Intelligence and
Data Science (ICCIDS 2018).
Keywords:Smart education; Artificial intelligence; Learning styles; Felder & Silverman; Kolb; Adaptive learning, Decision trees, Perceptrons
Keywords:Smart education; Artificial intelligence; Learning styles; Felder & Silverman; Kolb; Adaptive learning, Decision trees, Perceptrons

* Corresponding author. Tel.: +91-9899-630-501

E-mail address:author.
* Corresponding [email protected]
Tel.: +91-9899-630-501
E-mail address: [email protected]
1877-0509© 2018 The Authors. Published by Elsevier B.V.
Peer-review2018
1877-0509© under responsibility
The of thebyscientific
Authors. Published committee of the International Conference on Computational Intelligence and
Elsevier B.V.
Data Scienceunder
Peer-review (ICCIDS 2018). of the scientific committee of the International Conference on Computational Intelligence and
responsibility
Data Science (ICCIDS 2018).

1877-0509 © 2018 The Authors. Published by Elsevier Ltd.

This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/3.0/)
Peer-review under responsibility of the scientific committee of the International Conference on Computational Intelligence and Data Science
(ICCIDS 2018).
10.1016/j.procs.2018.05.095
 Richa Bajaj et al. / Procedia Computer Science 132 (2018) 834–842 835
Richa Bajaj, Vidushi Sharma/ Procedia Computer Science 00 (2018) 000–000

1. Introduction

Students learn in different ways. Some prefer facts, data and experiments whereas others prefer principles and
theories. Some prefer reading written material whereas others prefer problem solving. Learning management
systems so far have been developed with the philosophy of "one-size fits all" [15], as a result of which students tend
to get disoriented and the information overload results in reduced efficiency. Each student has his or her own
learning style.Determining a student’s learning style is a crucial step in making e-learning or traditional education
adaptive to students’ needs. There are multiple learning style models described in literature [4,5,6,9]. Some of the
prominent learning models are Felder & Silverman’s [4], Kolb’s [9], VARK [5]and Honey & Mumford [6]
model.These theories propose that all people can be classified according to their 'style' of learning, and provide
differing views on how the styles should be defined and categorized. These have been elaborated in Section 3 of this
paper.There are also numerous techniques described in literature that map a student’s behavioral attributes to a
particular learning style [3]. Artificial Intelligence(AI) approaches are regarded as valuable tools, as they have the
ability to develop and replicate the decision-making process adopted by people. There are various AI techniques that
have been used in adaptive educational systems. These include, but are not limited to, Fuzzy Logic, Decision Trees,
Neural Networks, Bayesian Networks, Genetic Algorithms and Hidden Markov Models [3].But there is no standard
approach created so far, to find out which is the most suitable learning theory and the most suitable artificial
intelligence method to applyfor a particular learning environment. Nor there is any software tool developed that
facilitates determining the learning style from data of students’ learning behavior. What is required is a tool that is
easily configurable, easily accessible and can be used in different learning environments, either traditional or e-
learning. Here, an artificial intelligence based system is developed that takes into consideration multiple learning
style models and multiple artificial intelligence techniques for determining students’ learning styles. This system can
be deployed both in e-learning and traditional educational environments to impart adaptive education.

2. Related work

Various AI methods that have been used earlier for providing adaption in learningare briefly reviewed here.
Fuzzy logic is an extension for the traditional set theoryas statements can be partial truths, lying in betweenabsolute
truth and absolute falsity. A multi-agent based student profiling system based on fuzzy logic has been given by
[17].By applying fuzzy logic, the content model, the student model, and the learning plan have been defined
formally. Neural networks comprise a large number of interconnected neurons which work together to process
information, similar to a biological neural network. These can be used to classify students. Previous studies have
shown the application of artificial neural networks in determining learning styles [1,11,14,16,18]. A decision tree is
a tree in which each branch node represents a choice between a number of alternatives, and each leaf node
represents a decision. In [10], decision trees have been used to provide personalized learning paths. A Bayesian
network is a directedacyclic graphin which nodes represent conceptsand edges indicatecause/effect dependencies
between concepts. In [13], the authors show the use of Bayesian Networks in the modeling of global personalized
learning process. In Hidden Markov Model, a set of discrete states are described with the probability matrix being
the determining factor of the transition between the states. Hidden Markov Models have been used in [8] to predict
student behavior and determine similarity between previous students and the current student. Genetic algorithms use
Darwin‘s concept of evolution, natural selection and survival of the fittest as their foundation. These have been
employed by [7] to construct an optimal learning path for eachlearner.

3. Existing Learning Theory Models

3.1 Felder & Silverman learning theory model

In the Felder and Silverman learning style model[4] there are four dimensions, each having two learning styles.
836 Richa Bajaj et al. / Procedia Computer Science 132 (2018) 834–842
Richa Bajaj, Vidushi Sharma/ Procedia Computer Science 00 (2018) 000–000

There is a total of eight learning styles that can generate sixteen combinations. These four dimensions are
active/reflective, sensing/intuitive, visual/verbal and sequential/global. Active learners learn best by working
actively, by applying the material, and by trying things out. In contrast, reflective learners prefer to think about and
reflect on the learning content. Learners with a sensing learning style use their sensory experiences to learn facts and
concrete learning material.In contrast, intuitive learners prefer to learn abstract material, such as theories and their
underlying meanings, with general principles. Visual learners remember best what they have seen e.g. pictures,
diagrams and flow-charts, whereas verbal learners learn from textual representations, which may be written or
spoken.Sequential learners learn in small incremental steps and therefore have a sequential learning progress.In
contrast, global learners use a global and a holistic thought process and learn by understanding the larger picture
first.

3.2 Kolb’s learning theory model

Kolb's learning theory [9] has four distinct learning styles, which are based on a four-stage learning cycle.The
four-stage learning cycle has: Concrete Experience - (CE), Reflective Observation - (RO), Abstract
Conceptualization - (AC) and Active Experimentation - (AE). Each of the fourlearning styles is a combination of
two cycle states. Diverging (CE/RO) learners prefer to watch rather than do, thus tending to gather information and
use their imagination to solve problems. Assimilating (RO/AC) learners prefer concise, logical approach. Ideas and
concepts are more important than people for them. Converging (AC/AE) learners solve problems and use their
learning to find solutions to practical problems. They prefer technical tasks, and are less concerned with people
oriented activities.Accommodating (AE/CE) learners are hands-on with tasks, and rely on their intuition rather than
logic. Theytend to use other’s analysis. They prefer a practical and experience oriented approach in learning.

3.3 Honey and Mumford learning theory model

Honey and Mumford learning styles were developed by Peter Honey and Alan Mumford [6]. Their work is
inspired from Kolb’s learning model. The four learning styles are activists, theorists, pragmatists and reflectors.
Activists areindividuals who learn by doing.The learning activities can be brainstorming, problem solving, group
discussion, puzzles, competitions or role-play. Theorists learners require models, ideas and theoriesto participate in
the learning process. Their learning activities include models, statistics, storiesand they applyconcepts
theoretically.Pragmatistshave the ability to put their learning into practice in reality.They learn better by applying
learning in case studies, in problem solving and in discussions. Reflectorslearn by watching, thinking and reflecting
on what happened. They like self-analysis and personality questionnaires, observation of activities, feedback from
others and interviews.

3.4 VARK learning theory model

VARK stands for Visual, Aural, Read/write, and Kinesthetic styles that are used for learning information. This
model was suggested by Fleming and Mills and is based on experiences of students and teachers [5].Visual learning
style includes maps, diagrams, charts, graphs, flow chartsand symbols that people use to represent information as a
replacement of words. Aural learning style describes a preference for information, which isheard or spoken.Learners
who have this learning style learn best from lectures, group discussion, radio, phones, speaking, web-chat and
talking about concepts. Those having Read/Write learning style prefer information displayed as words. It
emphasizes text-based input and output, reading and writing manuals, reports, essays and assignments. Kinesthetic
learning style refers to perceptual preference for experience and practice, which may be simulated or real. It includes
demonstrations, simulations, videos, movies and case studies.

4. Comparative Analysis of Models

Comparative analysis of models is done by implementing those models using artificial intelligence techniques.
These models have been implemented in Java 1.8 using NetBeans IDE 8.2 on a Windows 7 64-bit machine. Swing
 Richa Bajaj et al. / Procedia Computer Science 132 (2018) 834–842 837
Richa Bajaj, Vidushi Sharma/ Procedia Computer Science 00 (2018) 000–000

UI components have been used to design the GUI (graphical user interface).A software system for artificial
intelligence based adaptive learning has been developed here. Two learning models, Felder & Silverman [4] and
Kolb [9] have been simulated using the developed software, using two Artificial Intelligence techniques: Multilayer
Perceptron and Decision Trees.

The software system has menu systems for configuring student attributes that are to be considered for
determining learning styles. The screenshots for student attributes master for Felder & Silverman and Kolb models
are shown in Fig 1 and Fig 2 respectively. These screens show the attribute master, a superset of possible attributes
that have been presently built into the system. Either all or a subset of these may be selected for a particular
simulation run, depending on the attributes available in a particular learning environment. A typical execution is
depicted in Fig 3, showing the generation of model structure and its performance.

Analysis of model structures after carrying out training with sample data has been shown in Table 1. The first
column specifies the number of student attributes that were selected for model generation, here six for Felder &
Silverman and eight for Kolb’s model.

Fig. 1:Attribute master for Felder & Silverman model (a subset of attributes relevant to the learning environment are selected)

Student attributes for Felder & Silverman that have been selected for this particular simulation run are:
838 Richa Bajaj et al. / Procedia Computer Science 132 (2018) 834–842
Richa Bajaj, Vidushi Sharma/ Procedia Computer Science 00 (2018) 000–000

 Forum visits count (low/high): Number of visits to online forums is low or high
 Content visits count (low/high): Number of visits to online content is low or high
 SAT Factual visits count (low/high): Number of visits to factual type of questions in scholastic aptitude
tests is low or high
 SAT Abstract visits count (low/high): Number of visits to abstract type of questions in scholastic aptitude
tests is low or high.
 Course figures visits count(low/high): Number of visits to figures within online content is low or high;
 Course text visits count(low/high): Number of visits to text within online content is low or high

Fig. 2:Attribute master for Kolb’s model (all attributes are selected)

The student attributes used for the Kolb’s simulation are:

 Performance in brainstorming sessions (low/high);

 Prefer working in groups (low/high);
 Focus on ideas as compared to people (low/high);
 Attraction to logical theories as compared to practical ideas (low/high);
 Attraction to technical tasks rather than social issues (low/high);
 Finding practical uses for ideas and theories (low/high);
 Hands on approach, sets targets, active field work (low/high);
 Rely on intuition or others analysis rather than logic or own analysis (low/high).

The tool can also be used to specify the number of learning styles, into which the classification is to be carried
out. It is six for Felder & Silverman model and all four for Kolb’s model. For Felder & Silverman model, students’
classification is to be carried out into active/reflective, sensing/intuitive or visual/verbal categories. For Kolb’s
model, students’ classification is to be carried out into diverging, assimilating, converging and accommodating
 Richa Bajaj et al. / Procedia Computer Science 132 (2018) 834–842 839
Richa Bajaj, Vidushi Sharma/ Procedia Computer Science 00 (2018) 000–000

categories. This is depicted in the second column in Table 1. The third column in Table 1 shows the size of sample
data that has been used to generate the model structures. These can also vary based on the learning environment.
The tool generates neural network structures and decision trees each for Felder & Silverman and Kolb models. The
fourth column in Table 1 shows the number of nodes generated in the model. For the neural network, the multilayer
perceptron models have two layers each. The output layer has six neurons for the Felder & Silverman model and
four neurons for the Kolb’s model, based on the number of classification categories. Similarly, the decision trees
that have been generated have six leaf nodes for Felder & Silverman model; and four leaf nodes for Kolb’s model.
Comparisons of model performances have been shown in Table 2. Neural networks have exhibited higher
performance as compared to decision trees. This can be seen from the Kappa statistics [2] values. The multilayer
perceptron model’s performance is slightly better for Kolb’s as compared to Felder & Silverman’s as it has lower
root mean squared error value. Also, the decision tree model’s performance is comparable for both Felder &
Silverman and Kolb’s models. Kappa statistics value is higher for Kolb’s model; with a slightly higher root mean
squared error value as compared to the decision tree for Felder & Silverman model. The data show here is for a
typical execution of the simulation. A simulation first generates a model structure and then measures the model’s
performance.

Fig. 3:A typical execution showing generation of a neural network model and its performance for Felder & Silverman learning theory

Table 1:Comparison of generated model structures

Model / AI Method Number of Number of Size of student Number of
student attributes learning styles sample data nodes in model
(classification
classes)
Felder Silverman/Multilayer Perceptron 6 6 22 12
Felder Silverman/Decision Tree 6 6 22 11
Kolb/Multilayer Perceptron 8 4 12 10
Kolb/Decision Tree 8 4 12 7
840 Richa Bajaj et al. / Procedia Computer Science 132 (2018) 834–842
Richa Bajaj, Vidushi Sharma/ Procedia Computer Science 00 (2018) 000–000

Table 2:Comparison of model performance

Model / AI Method Correctly Incorrectly Kappa statistics Root mean
classified classified squared error
instances instances
Felder Silverman/Multilayer Perceptron 22 0 1 0.0393
Felder Silverman/Decision Tree 15 7 0.6051 0.2611
Kolb/Multilayer Perceptron 12 0 1 0.0351
Kolb/Decision Tree 9 3 0.6667 0.2887

5. Framework of a Smart Education Model

A framework of a tool for smart education is proposed here. Smart education is about providing personalized
learning, anywhere and anytime. To make this realistic and widely available, development of a software system for
artificial intelligence based determination of learning styles, is proposed here. The determined learning styles can be
used to make learning content adaptive and push different content to different students based on their respective
learning styles. As of now we have shown simulations with two artificial intelligence techniques and two learning
style theories. The software system is being developed to include more learning theories such as Honey and
Mumford [6] and VARK’s [5]; and also more artificial intelligence methods such as Fuzzy Logic, Genetic
Algorithms, Bayesian Networks and Hidden Markov Models. A review of various methods that can be applied has
been given in [3]. Also, it is expected that wider variations in data set may be taken in future, to include different
leaning environments. These learning environments could be traditional schools, higher education colleges,
institutes in villages or institutes in cities. The variations in data could be due to different courses such as arts,
science and engineering; or due to e-learning environments with text, with multimedia, with online aptitude tests or
with voice based interfaces. The rich set of student attributes for each learning theory model would facilitate
selection of the most appropriate learning theory model for a particular environment.

Fig. 4:Framework for smart education showing students interacting with a virtual teacher on a cloud environment. It depicts determination of
learning styles using the most suitable learning theory model and artificial intelligence technique.
 Richa Bajaj et al. / Procedia Computer Science 132 (2018) 834–842 841
Richa Bajaj, Vidushi Sharma/ Procedia Computer Science 00 (2018) 000–000

The framework suggests a multi-step process for delivering personalized smart education. This has been described
below and also depicted in Fig 4.
 Identification of student attributes for a particular learning environment where adaptive learning is to be
provided
 Based on available student attributes, selection of one or more learning theory models that can be possibly
applied. Learning models that do not have sufficient available attributes would get eliminated.
 In case more than one learning theory model is applicable for an educational environment, the tool would
help in determination of the most suitable model, based on performance of different models.
 The tool would also facilitate determination of the most suitable artificial intelligence method that should be
used to build the final classification model. This would be based on comparison of model performances. The
model with highest Kappa statistics value and least mean squared error value would be the most preferred
one.
 Once one or more models have been shortlisted and trained for a particular student environment, these may
be used to classify students and determine their learning styles. Learning styles can be mapped to learning
content and learning paths to deliver personalized education.

Smart education is about taking learning outside the traditional classrooms; and is an activity that can be done
anywhere and anytime. Internet enabled tablets to browse personalized learning content that could be text, images or
multimedia, internet enabled watches to listen to recorded lectures are some of the devices that can be used. To
enable this,use of cloud technologies is suggested here. The software tool being developed can be hosted on a cloud
environment for easy access to worldwide audience, without limitations of scalability. Sample learning content
would be required to be generated to monitor student behavior and determine the available student attributes. Instead
of an actual classroom, a virtual classroom and a virtual teacher may be used. It is suggested that natural language
processing [12] APIs, voice to text and text to voice APIs, may be used in future to generate sample learning
content; and simulate a student-teacher interaction, with virtual teacher agents. For this, cloud based natural
language processing technologies such Microsoft Luis, Amazon Lex or IBM Watson may be used. This would
enable students to interact with a virtual teacher by just speaking to the system; as if speaking to a human teacher.

So far, adaptive education has been primarily a research area with scattered implementations of some learning
models. With the framework for smart education being suggested here, it is proposed that adaptive education be
offered as a cloud based service; so that numerous traditional schools, colleges and e-learning platforms can easily
access this service and deliver personalized education to their students.

6. Conclusions

The above framework for smart education, which has been developed here, would help in the future in making
adaptive education easily available to a wide audience of students, across different cultural backgrounds,
geographies and modes of education, traditional or eLearning. The framework provides a collection of numerous
student learning attributes that can be tracked, based on which personalized learning can be provided. This readily
available collection, within a single tool, would make it easy to determine which student learning attributes can be
selected for a particular learning environment. This would facilitate the implementer to narrow down on possible
learning theory models that can be applied to impart adaptive learning. This is the first ever framework that has been
developed to compare multiple learning theories; and compare artificial intelligence based classification
techniques,based on performance of developed models.These models are developed dynamically, within the same
tool; and statistical evaluation helps determine the most suitable model, for implementation in a given learning
environment. The selected learning theory and artificial intelligence method can then be used to determine the
learning styles of students.The framework suggests for a virtual teacher, to be hosted on a cloud environment; that
interacts with learners in a scalable manner using natural language processing APIs, to dynamically determine their
learning styles. The determined learning styles of students can be subsequently used by various learning content
providers, either traditional schools or e-learning portals, to provide adaptive education.
 Richa Bajaj, Vidushi Sharma/ Procedia Computer Science 00 (2018) 000–000

842 Richa Bajaj et al. / Procedia Computer Science 132 (2018) 834–842

References

[1] Bernard J, Chang TW, Popescu E, Graf S. Using artificial neural networks to identify learning styles. InInternational Conference on
Artificial Intelligence in Education 2015 Jun 21 (pp. 541-544).Springer, Cham.
[2] Carletta J. Assessing agreement on classification tasks: the kappa statistic. Computational linguistics. 1996 Jun 1;22(2):249-54.
[3] Colchester K, Hagras H, Alghazzawi D, Aldabbagh G. A survey of artificial intelligence techniques employed for adaptive educational
systems within E-learning platforms. Journal of Artificial Intelligence and Soft Computing Research. 2017 Jan 1;7(1):47-64.
[4] Felder RM, Silverman LK. Learning and teaching styles in engineering education. Engineering education. 1988 Apr 1;78(7):674 -81.
[5] Fleming ND. Teaching and learning styles: VARK strategies. IGI Global; 2001.
[6] Honey P, Mumford A. Styles of learning. Gower Handb. Manag. Dev. 1994;101:101-11.
[7] Huang MJ, Huang HS, Chen MY. Constructing a personalized e-learning system based on genetic algorithm and case-based reasoning
approach. Expert Systems with Applications. 2007 Oct 1;33(3):551-64.
[8] Huang X, Yong J, Li J, Gao J. Prediction of student actions using weighted Markov models. InIT in Medicine and Education, 200 8.ITME
2008. IEEE International Symposium on 2008 Dec 12 (pp. 154-159). IEEE.
[9] Kolb AY, Kolb DA. Learning styles and learning spaces: Enhancing experiential learning in higher education. Academy of manage ment
learning & education. 2005 Jun 1;4(2):193-212.
[10] Lin CF, Yeh YC, Hung YH, Chang RI. Data mining for providing a personalized learning path in creativity: An application of decision trees.
Computers & Education. 2013 Oct 1;68:199-210.
[11]Lo JJ, Shu PC. Identification of learning styles online by observing learners’ browsing behaviour through a neural network. British Journal
of Educational Technology. 2005 Jan 1;36(1):43-55.
[12] Manning CD, Schütze H. Foundations of statistical natural language processing. MIT press; 1999.
[13] Moreno F, Carreras A, Moreno M, Royo ER. Using bayesian networks in the global adaptive e-learning process.EUNIS 2005. 2005:1-4.
[14] Mota J. Using learning styles and neural networks as an approach to elearning content and layout adaptation.InDoctoral Sympos ium on
Informatics Engineering 2008.
[15] Oliver R, Harper B, Reeves T, Strijker A, van Westhuizen D. Learning management systems: One size fits all?. InWorld Conference on
Educational Multimedia, Hypermedia & Telecommunications 2002 2002 Jun 24.
[16] Villaverde JE, Godoy D, Amandi A. Learning styles' recognition in e ‐learning environments with feed ‐forward neural networks. Journal of
Computer Assisted Learning. 2006 Jun 1;22(3):197-206.
[17] Xu D, Wang H, Su K. Intelligent student profiling with fuzzy models. InSystem Sciences, 2002.HICSS.Proceedings of the 35th Annual
Hawaii International Conference on 2002 Jan 7 (pp. 8-pp).IEEE.
[18] Zatarain-Cabada R, Barrón-Estrada ML, Angulo VP, García AJ, García CA. A learning social network with recognition of learning styles
using neural networks. InMexican Conference on Pattern Recognition 2010 Sep 27 (pp. 199-209).Springer, Berlin, Heidelberg.