Original article

doi: 10.1111/j.1365-2729.2007.00262.x

An enhanced Bayesian model to detect students learning styles in Web-based courses

P. Garca, S. Schiafno & A. Amandi
ISISTAN Research Institute, Facultad de Ciencias Exactas, Universidad Nacional del Centro de la Pcia de Buenos Aires, Tandil, Buenos Aires, Argentina and CONICET, Consejo Nacional de Investigaciones Cientcas y Tcnicas, Buenos Aires, Argentina

Abstract

Students acquire and process information in different ways depending on their learning styles. To be effective, Web-based courses should guarantee that all the students learn despite their different learning styles. To achieve this goal, we have to detect how students learn: reecting or acting; steadily or in ts and starts; intuitively or sensitively. In a previous work, we have presented an approach that uses Bayesian networks to detect a students learning style in Webbased courses. In this work, we present an enhanced Bayesian model designed after the analysis of the results obtained when evaluating the approach in the context of an Articial Intelligence course. We evaluated the precision of our Bayesian approach to infer students learning styles from the observation of their actions with a Web-based education system during three semesters. We show how the results from one semester enabled us to adjust our initial model and helped teachers improve the content of the course for the following semester, enhancing in this way students learning process. We obtained higher precision values when inferring the learning styles with the enhanced model. Bayesian networks, e-learning, learning styles, student modeling.

Keywords

Introduction

Students learn in many different ways (Felder & Silverman 1988): by seeing and hearing; reecting and acting; reasoning logically and intuitively; memorizing and visualizing; drawing analogies and building mathematical models; steadily and in ts and starts. Similarly, teachers teach with different methods. Some teachers lecture, others discuss or demonstrate; some emphasize memory while others understanding. Studies have shown that greater learning may occur when the teaching style matches the students learning styles than when they are mismatched (Allinson & Hayes 1996; Felder & Brent 2005).

Accepted: 13 September 2007 Correspondence: Silvia Schiafno, ISISTAN, Campus Universitario, Paraje Arroyo Seco, Tandil, Buenos Aires, Argentina. Email: sschia@ exa.unicen.edu.ar

In the context of Web-based courses, the teaching styles are characterized by: the content of courses, that is, texts, examples, exams and exercises proposed; how information is presented to students; the interaction mechanisms among students and between students and teachers, such as chat rooms, forums and other collaborative or groupware tools; and the order in which contents are organized and presented within a course. One of the most desired characteristics of a Web-based education system is that of being adaptive and personalized (Brusilovsky & Peylo 2003), as it has to be used by a wide variety of students with different skills, preferences and learning styles. To achieve this goal, it is vital to discover each students learning style and adapt the courses or assist the student according to it. The system or an intelligent tutor could then assist students by suggesting to them courses of actions that correspond to the students learning style, or by suggesting courses of actions that supplement the negative
305

2007 The Authors. Journal compilation 2007 Blackwell Publishing Ltd

Journal of Computer Assisted Learning (2008), 24, 305315

306

P. Garca et al.

aspects of a certain style depending on the approach adopted. In a previous work (Garcia et al. 2007), we proposed the use of Bayesian networks (BN) to represent and automatically detect students learning styles in a Web-based education system. This Articial Intelligence technique enables us to model both quantitative and qualitative information about students behaviour with a Web-based educational system. Different students actions that are observed while the student works with the learning system are modeled in the BN, such as the type of material preferred or the number of exercises done. More importantly, the inuence of these actions on the different dimensions that compose a learning style is also modeled. The strength of these dependences is specied by conditional probability values that indicate, for example, that if the type of material preferred by a student is abstract, then it is quite probable that the student reasons intuitively. The information used to build the Bayesian model is obtained by analysing students log les that contain records of the tasks carried out by the students in the system. We evaluated the proposed approach with students of an Articial Intelligence course and we obtained promising results. However, we also detected some aws not only in the proposed model but also in the design of the underlying Web-based course that yield to low precision in the detection capability of our technique. In this article, we present the enhanced Bayesian model and we describe the different experiments we carried out during three semesters to evaluate our proposal. We show how the results from one semester helped teachers improve the content of the course for the following semester, enhancing in this way students learning. We obtained higher precision values when inferring students learning styles with the enhanced Bayesian model, than with the initial model. This article is organized as follows. In the next section, we present the new Bayesian model to infer students learning styles. After which, we present the experiments carried out to evaluate the proposed model. Also we present some related works. Finally, we present our conclusions and future work.
Detecting learning styles with BN

Table 1. Dimensions of Felders learning styles (those marked in italics are the ones considered in this work). Perception Input Processing Understanding Organization Sensitive Intuitive Visual Verbal Active Reective Sequential Global Inductive Deductive

learning styles automatically based on the observation of their actions. The initial proposal is described in detail in the study by Garcia et al. (2007). This model showed some difculties at detecting some of the dimensions of students learning styles. In the following subsections we describe a new model that incorporates new variables to enable a more accurate detection of these styles.

Learning styles

Our approach uses BN both to model students behaviours while taking a Web-based course and to infer their

A learning style model classies students according to where they t on a number of scales belonging to the ways in which they receive and process information. Several models and frameworks for learning styles (Kolb 1984; Felder & Silverman 1988; Honey & Mumford 1992; Litzinger & Osif 1993) have been proposed. Although they have been critically analysed by some authors (Cofeld et al. 2004), the use of learning styles for experimental research in Web-based educational systems has demonstrated that providing material according to students learning styles can enhance students learning (Walters et al. 2000; Pea et al. 2002; Paredes & Rodriguez 2004), and that cognitive styles are linked to quantitative differences in both navigation behaviour and learning performance (Dufresne & Turcotte 1997; Ford & Chen 2000; Chen et al. 2005). In this work we use the model proposed by Felder and Silverman for engineering students (Felder & Silverman 1988). This model comprises 32 learning styles. Table 1 shows the dimensions of the learning styles in this model. We chose this model after analysing several learning styles, models and frameworks. First, because it was designed for engineering students and second, because we observed that for students of exact
2007 The Authors. Journal compilation 2007 Blackwell Publishing Ltd

Students learning styles in Web-based courses

307

sciences (math students, computer science students), providing material and teaching according to students perception dimension values helped them improve their understanding in regular classes. Thus, we decided to take this experience from classrooms to e-learning. According to this model, sensors like facts, data and experimentation; intuitive students prefer principles and theories. Sensors are patient with detail but do not like complications; intuitive students are bored by detail and welcome complications. Visual learners remember best what they see: pictures, diagrams, time lines, lms and demonstrations. Verbal learners remember much of what they hear or read, and more of what they hear or read and say. Active learners do not learn much in situations that require them to be passive, and reective learners do not learn much in situations that provide no opportunity to think about the information being presented. Active learners work well in groups; reective learners work better by themselves or with at most one other person. Sequential learners follow linear reasoning processes when solving problems; global learners make intuitive leaps and may be unable to explain how they came up with solutions. Sequential learners can work with material when they understand it partially or supercially, while global learners may have great difculty doing so. Induction is a reasoning progression that proceeds from particulars to generalities. Deduction proceeds in the opposite direction. Induction is the natural human learning style. Experiments have proved that most engineering students are inductive learners (Felder & Silverman 1988). Thus, this dimension is not considered and 16 learning styles remain. In this work, we consider only three dimensions of Felders framework, namely perception, processing and understanding. They are marked in italics in Table 1. We discarded the input dimension because we are currently not considering videos or simulations as part of the Web courses.
Bayesian networks

Bayesian networks, also known as causal networks, constitute an Articial Intelligence technique widely used to model uncertain domains. Nodes in a BN represent variables of interest in such domains and arcs connecting the nodes represent causal relationships between these variables (Jensen 2001). BN enable us to
2007 The Authors. Journal compilation 2007 Blackwell Publishing Ltd

model both qualitative and quantitative information. Qualitative information is given by the network structure, that is, the nodes, the values of the nodes, and the arcs that connect them. In this work, the nodes in the network represent the different dimensions of learning styles and those students actions we observe to determine them. The arcs in the network represent the relationships between the behaviours observed while a student uses the system and the learning style dimensions studied. Each node in a BN can have different states, corresponding to the various values that the variable can take. Variables can have two (true and false, for example) or more values. The arcs in a BN are directed. They go from the cause (parent node) to the effect (child node). Nodes without parents are known as independent nodes. In our work, the direction of arcs indicates that a given student behaviour with the learning system determines a certain value in a learning style dimension. In other words, they indicate that the learning style depends on, or is determined by, the actions the student carried out. On the other hand, quantitative information in a BN is expressed through the strength of the relationships between nodes, which are given by probability tables. For each node, a probability table species the probability of each possible state of the node given each possible combination of states of its parents. These tables are known as conditional probability tables (CPT). Tables for independent nodes just contain unconditional or simple probabilities. The values of probabilities can be obtained by different means, depending on the problem we are modeling. They can be specied by an expert in the domain considered; they can be obtained from statistical information; they can be calculated from a set of data using some learning algorithm or from observation of facts. For example, in the simple BN shown in Fig 1, behaviour_x and behaviour_y are independent nodes representing actions in a Web-based system, and style_dimension depends on the values that behaviour_xand behaviour_ytake. The direction of the arcs indicate this dependence. The gure also shows the probability tables that dene the quantitative information of the BN, supposing that behaviour_x and behaviour_y can take the values observed and not observed, and style_dimension can take the values style_one and style_two. In this example, the probability values indicate that behaviour_x was observed

308

P. Garca et al.

behaviour_x=observed 0.7

behaviour_x=not observed 0.3

behaviour_x

behaviour_y

behaviour_y=observed 0.6

behaviour_y=not observed 0.4

style_dimension

behaviour_x=observed behaviour_x=observed behaviour_x=not_ behaviour_x=not_ behaviour_y=observed behaviour_y=not_ observed observed observed behaviour_y=observed behaviour_y=not_ observed style_dimension=style_one 0.5 0.3 0.8 0.5 style_dimension=style_two 0.5 0.7 0.2 0.5 P(style_dimension/ behaviour_x, behaviour_y)

Fig 1 A simple Bayesian network.

70% of the times the student used the system and behaviour_y was observed 60% of the times. In the style_dimension CPT, we specify the probability of the node having the values style_one and style_two given each possible combination of the values of behaviour_x and behaviour_y. The CPT indicates, for example, that if behaviour_x is observed and behaviour_y is not observed in a set of a students actions, then the probability of the student having style_two is 70%. In addition to BNs simplicity to model different domains in terms of variables and the relationships between them, the mathematical model underlying BN makes them a sound technique and enables us to make inferences about the value of a certain node given the observation of values in other nodes in the network. The mathematical model underlying BN is Bayes theorem, which is shown in equation (1). Bayes theorem yields the conditional probability distribution of a variable A, assuming we know: information about another variable B in terms of the conditional probability distribution of B given A, and the probability distribution of A alone. Equation (1) reads: the probability of A given B equals the probability of B given A times the probability of A, divided by the probability of B.

behaviour (behaviour_x for example) and A can stand for a style dimension. Then, considering information about a set of students actions, P(A) is the probability of a student having a certain value in a style dimension (this probability can be obtained by counting the number of students having this style dimension over the total number of students observed). P(B) is the probability of a student showing a particular behaviour (it can be obtained by counting number of occurrences of actions in the log le). P(B|A) is the probability of a student showing a particular behaviour given that the student has a certain learning style (this probability can be also obtained from the information in the log le). Then, we could infer the value of the node style_dimension given evidences of a xed value in the node behaviour_x, that is P(A|B). For more complex networks, generalized versions of the Bayes theorem are used when trying to infer the probabilities of nodes given evidence of the values of other nodes (Pearl 1988). The use of BN implies two steps: building a BN that models the problem or domain; and using the model built to make inferences about this domain. In the following subsection, we describe how we use this technique to infer students learning styles.
Using BN to infer learning styles

P ( A B) =

P ( B A) P ( A) P ( B)

(1)

Bayes theorem is the key of Bayesian inference mechanisms. Through these mechanisms, we can infer the value of any node in the network given evidence of the values taken by other nodes. In the example of Fig 1, B in equation (1) could represent a given student

To use BN we must, rst, build a network that contains the variables of interest and the relationships between these variables, and second, assign the probability distribution to each node in the graph in order to indicate the strength of the relationships previously modeled.
2007 The Authors. Journal compilation 2007 Blackwell Publishing Ltd

Students learning styles in Web-based courses

309

Table 2. Behaviours observed to detect learning styles. Perception Time dedicated to exam revision Time it takes the student to nish and submit the exam Number of changes in exam answers Type of reading material preferred (abstract, concrete) Number of exercises done Number of examples read Participation in forums Use of chat and mail systems Participation in collaborative/group tasks Pattern of access to information Exam results

Processing

Understanding

In our domain, variables represent the different dimensions of Felders learning styles and the behaviours that determine each of these dimensions. As said before, we consider only three dimensions of Felders framework, namely perception, processing and understanding. We model each dimension with a variable in the BN. The values these variables can take are sensory/intuitive, active/reective, and sequential/ global, respectively. The behaviours we analyse to determine the different dimensions of learning styles are summarized in Table 2. This information is obtained by analysing the data recorded in a students log le. With respect to the Perception dimension we can say that, according to Felder, a student who does not revise his/her exercises or exams is likely to be intuitive. On the other hand, a student who carefully checks the examinanitons or exercises is generally sensory. A student who reads or accesses various examples of a given topic is more sensory than one who reads just one or two. As regards the type of reading material the student prefers, a sensory learner prefers concrete (application-oriented) material, while an intuitive learner usually likes abstract or theoretical texts. To detect whether the student prefers to work things out alone (reectively) or in groups (actively), we analyse his/her participation in forums, chats and mail systems. As regards forums, we analyse whether the student begins a discussion, replies to a message or just reads the messages posted by other students. The frequency of this participation is also important. The participation in chat and mails can give us some information, but it is not as relevant as the one we can obtain with a forum access log. To
2007 The Authors. Journal compilation 2007 Blackwell Publishing Ltd

enable students to work collaboratively, our Web-based education system (named SAVER) provides a collaborative facility. Given a certain problem-solving activity, students use this tool to propose solutions to the problem, to make a counterproposal to a proposed solution and to read the solutions available thus far. They can also send messages to other members of the group or read messages posted by others. In addition, the system logs the participation of each student in the group activity. Finally, to determine how students understand, we analyse access patterns to information, which are recorded in students log les. If the student jumps through the course contents, we can say that he/she does not learn sequentially but in ts and starts. The results the student gets in the exams while he/she is jumping over the contents give us an indication of his/her understanding style. If the student gets a high mark in a topic despite having not read a previous topic, we can conclude that the student does not learn sequentially. The dependencies between behaviours and learning styles that we mentioned before are encoded in the network structure. The resultant network is shown in Fig 2. The following sentences describe in detail the different states the independent nodes can take: forum: posts messages; replies messages; reads messages; no participation. chat: participates; listens; no participation. mail: uses; does not use. tasks: makes proposal for group task; makes counterproposal; reads proposal. messages: sends message; reads message (within group task). participation: participates; no participation. information access: in ts and starts; continuous. reading material: concrete; abstract. exam revision (considered in relation to the time assigned to the exam): less than 10%; between 10% and 20%; more than 20%. exam delivery time (considered in relation to the time assigned to the exam): less than 50%; between 50% and 75%; more than 75%. exercises (in relation to the amount of exercises proposed): many (more than 75%); few (between 25% and 75%); none.

310

P. Garca et al.

Fig 2 Bayesian network modeling a students learning style.

answer changes (in relation to the number of questions or items in the exam): many (more than 50%); few (between 20% and 50%); none. access to examples (in relation to the number of examples proposed): many (more than 75%); few (between 25% and 75%); none. exam results: high (more than 7 in a 110 scale); medium (between 4 and 7); low (below 4). The nal step in constructing a BN is to determine the probability values, that is, the strength of relationships. The model is completed by establishing the probability values associated with each node in the network. Our model contains 18 probability tables. With respect to the quantitative aspect of the model, the major change in the enhanced BN is the reformulation of the probability function that corresponds to the Processing variable. As new factors determine this dimension, the corresponding probability function has been redened. As an example, Fig 3a shows the probability values obtained for a certain student for the Forum node. The third cell of the second column in the Forumtable indicates that 50% of the times the student used the application he/she posted messages to the forum. Fig 3b shows the CPT for the Understanding node. For example, the second cell in the rst (numbered) column indicates that if the student reads in ts and starts and he/she gets high marks in the exams, the probability that this student is a global learner is 100%.

Fig 3 Probability table for (a) Forum and (b) CPT for Understanding given Exam Results and Information Access.

Initially, probability values for independent nodes are assigned equal values. Then, the values are updated as the system gathers information about the student behaviour. The probability functions attached to the independent nodes are adjusted to represent the new observations or experiences (Olesen et al. 1992). Consequently, the Bayesian model is continuously updated while new information about the students interaction with the system is obtained. On the other hand, the probability values contained in the different CPT were obtained via a combination of expert knowledge and experimental results. The expert knowledge was obtained from Felders denition of learning styles. We
2007 The Authors. Journal compilation 2007 Blackwell Publishing Ltd

Students learning styles in Web-based courses

311

took into account the inuence of the different factors analysed on the dimensions of the learning styles. To determine the values experimentally, we gave a set of 50 Computer Science Engineering students the ILS (Index of Learning Styles) questionnaire.1 This questionnaire determines a students learning style through a set of 44 questions. We considered these learning styles as correct. Then, we let these students use the education system to learn a topic of an Articial Intelligence course, and recorded their interactions with the system.2 Information about their recorded behaviour was used to determine the conditional parameters of the BN, in combination with the expert knowledge. It is worth noting that the values in the CPT are equal for all students, as they represent the strength of relationships between different behaviours and learning styles. However, the values corresponding to independent nodes are different for different students as they represent the actions taken by a particular student and are obtained from each students log le. Once we have built the BN and we have determined the values of the probability tables, our goal is inferring the values of the nodes corresponding to the dimensions of a learning style given evidences of the students behaviour with the system. Thus, we obtain the probability values of the learning style node given the values of independent nodes. The learning style of the student is the one having the greatest probability value. For example, suppose that we want to determine whether the student learns sequentially or globally, we have to compute the probability p(Understanding = Sequential), that is p(Understanding = Sequential/ Information Access, Exam Results), and p(Understanding = Global), that is, p(Understanding = Global/Information Access, Exam Results). Then, the value of the dimension is the one with the highest probability. If the rst probability is 75.5% and the second 24.5%, then the student learns sequentially. These computations, which are based on Bayes theorem, are done by Bayesian inference mechanisms implemented by the software we use, namely Hugin.3
Experimental results

knowledge on the topics taught in the Web-based course. We considered a course unit or topic as the minimum observation unit. The results obtained in our experiments were computed by averaging the students behaviour in different units. Each unit is, in turn, divided into a set of topics. Each topic has reading material available that is categorized either as abstract or concrete. Students are presented with a number of examples for a certain topic, and they can optionally have access to more examples if they need them. Similarly, each unit has a set of exercises students can optionally do to test themselves. The system automatically marks these exercises. In most cases, there are no prerequisites within a course. That is, no units are mandatory within a course and no previous units are required to read a given topic. At the end of the course, students must submit a nal exam. Students were encouraged to do all their studying through the e-learning system. All the material was provided through this media. However, some students (3%) did not use the system properly and downloaded the material to print it and read it. The log les of these students were discarded in our experiments. We considered only the valid or useful logs.
Methodology

We evaluated our approach through an Articial Intelligence course taken by Systems Engineering students using an e-learning system named SAVER. The students involved in the experiment had no previous
2007 The Authors. Journal compilation 2007 Blackwell Publishing Ltd

To evaluate the precision of our approach, we compared the learning style detected by the BN model against the learning style obtained with the ILS questionnaire proposed by Felder, which is a tool to determine a students learning style. As we have mentioned before, in Felders model, each dimension of a learning style can take one of two possible values. However, in the ILS questionnaire, the author denes a scale that indicates, for example, how intuitive or sensitive a student is. This scale goes from 1 to 11 for each value of a given dimension. If a students score on the scale is 13, he is fairly well balanced on the two dimensions of that scale. If the score on a scale is 57, the student has a moderate preference for one dimension of the scale. If the score on a scale is 911, the student has a very strong preference for one dimension of the scale. Thus, in our experimental results, we considered three values for each dimension: the two extreme values, and a neutral value corresponding to the 13 scores. As described earlier, a value of neutral in a learning style dimension means that the student does not show a tendency towards one of

312

P. Garca et al.

the extreme values of the dimension considered and that he/she can learn or process information in both ways. The probability values obtained from the BN were also transformed to three values. A student showing a probability of 50%60% of having a certain value in a dimension was considered as neutral. Otherwise, the extreme value with the greatest probability value was considered. We assumed in our calculus that the learning style determined by the ILS questionnaire is correct. We used the formula in equation (2) to compute the precision of the technique. In this equation, Sim is 1 if the values obtained with the BN (LSBN) and ILS (LSILS) are equal, 0 if they are opposite, and 0.5 if one is neutral and the other an extreme value; and n is the number of students participating in the experiment.

most students made little use (or no use) of the chat, mail and forum facilities. The causes of this phenomenon were: the initial Web-based course did not promote the use of chat and forum facilities, and students did not normally use these technologies for studying. most students have read the whole course, that is, they read most of the theoretical material, they did most of the exercises proposed and they read most of the examples. most students did not skip units and read the material sequentially, independently of their learning styles. many students had little or no experience at working with Web-based courses.
Experiments with the enhanced Bayesian model

Precision =

sim ( LSBN , LSILS )

i =1

(2)

In the following sections, we describe the results obtained when evaluating the Bayesian model. The experiments were carried out during three editions of the Articial Intelligence course, that is, during three semesters.
Evaluation of the initial model

The rst experiment was carried out in 2005 (Semester 1) with 27 users with the initial model. The population of students with respect to their learning styles (according to ILS) was as follows. As regards the understanding dimension, 35% of the students were global, 15% sequential and 48% neutral. With respect to perception, 37% of the students were sensitive, 30% intuitive and 33% neutral. Finally, regarding the processing dimension, 30% of students were reexive, 33% neutral and 37% active. Using the formula in equation (2), we obtained a precision of 77% in the perception dimension, 63% in the understanding dimension, and 58% in the processing dimension. This last percentage was computed only considering the styles that could be obtained by the BN (52% of the students). According to the initial BN, the most popular processing style was reective. We further analysed the lack of information and the mismatches in this dimension by surveying the students who participated in the experiment. We found the following causes:

In the rst experiment, we discovered that we should encourage students to make use of facilities such as forums and chats in order to discover the active learners. Moreover, they should carry out collaborative tasks so that we could measure their participation in group activities. We also found that the inexperience of students at working with Web-based courses made them study differently from the way they usually do. Finally, we learned that to detect correctly the sequential/global dimension we have to observe students behaviour while working with big courses, where the size is measured in terms of numbers of units, examples and exercises. To enhance the Web-based course, and hence our model, we applied the following changes. We included: more units; more topics per unit; more examples and exercises; a group task as part of the evaluation of the course; a collaborative tool to carry out the task mentioned before.

The aim of the rst three changes was obtaining a more complete higher-quality course that would enable us to obtain more precise results in the understanding dimension. The incorporation of a group task as part of the evaluation of the course is a way of forcing students to use collaborative tools, and a way of evaluating more precisely the processing dimension. As said before, this
2007 The Authors. Journal compilation 2007 Blackwell Publishing Ltd

Students learning styles in Web-based courses

313

Table 3. Experimental results summary. Experiment No. of units No. of topics No. of examples 6 9 9 No. of exercises 5 10 10 Processing precision (%) 58 56 66 Understanding precision (%) 63 70 72 Perception precision (%) 77 77 80

First Second Third

1 7 7

7 14 14

leads to the incorporation of new variables in the Bayesian model and new probability functions in the corresponding CPT. We carried out an experiment using the enhanced course and the new BN model during the rst semester of 2006 with 49 students. The distribution of their learning styles is as follows. With respect to the perception dimension, 8% of students were intuitive, 55% neutral and 37% sensitive. As regards processing, 6% of the students were reexive, 37% active and 57% neutral. Finally, considering the understanding dimension, 16% were sequential, 64% neutral and 20% global. In this second experiment, we obtained a precision of 77% in the perception dimension, 70% in the understanding dimension and 56% in the processing dimension (computed considering 98% of the students). Then, we conducted a third experiment during the second semester of 2006 with 30 students. The population of students learning styles was the following. Considering the perception dimension, 33% of the students were sensitive, 57% neutral and 10% intuitive. With respect to the understanding dimension, 63% of the students were global learners, 20% neutral and 17% were sequential. Finally, considering the processing dimension, 7% of students were reexive, 63% neutral and 30% active. In this third experiment, we obtained a precision of 80% in the perception dimension, 72% in the understanding dimension and 66% in the processing dimension (computed considering 73% of the students).
Discussion

Table 3 shows the characteristics of the Articial Intelligence course in terms of number of units, number of topics, number of exercises and number of examples. It also shows the precision obtained for the different dimensions. We can observe that a more organized, more complete course was used during the last two experiments. This factor inuenced the way in which
2007 The Authors. Journal compilation 2007 Blackwell Publishing Ltd

students read the course and used the material provided. It also enabled us to record more information about students behaviour. Thus, the number of experiences registered increased in the last two experiments. This increment leads to a higher precision in the detection of the learning style to which students t in. As we explained before, if we have few experiences, they have a great impact in the calculus of probabilities, and hence, the probability values learned by the BN are not representative. Consequently, the style assigned to a student might be wrong. With respect to the precision of the BN to infer the different dimensions, we observe an increase of 3% in the Perception dimension. The precision values for this dimension are around 80% in all the experiments. As regards understanding, the precision values are about 70% in the last two experiments. We observe an increase of about 7% with respect to the initial model. The inclusion of more exercises, examples and topics within the course enabled the observation of a higher number of students actions for the detection of these two dimensions. Finally, we observe that in the Processing dimension, we observe an increase of 8% in the precision. Although we did not get a considerable increase in the prediction precision in this dimension, the number of students tested and the number of experiences recorded are higher in the rst and second experiments. In the rst experiment, the number of students whose processing dimension could be determined automatically was only 14, while in the second experiment, it was 48, and in the third experiment, it was 22. We are aware that the number of students that took part in the experiment might be small for statistical validity. We will carry out more experiments with bigger groups of students in the future. In addition, for future experiments, we will revise the probability values in the CPT affecting the Processing dimension, as we still experience some difculties at detecting the active and global learners.

314

P. Garca et al.

Related work

Numerous works have addressed the problem of student modeling (see (Brusilovsky & Peylo 2003) for a review). These works can be categorized according to different characteristics, such as the content of the student model, the type of student being modeled, how the student model is updated, what the model is used for, among others. Our work can be placed among those modeling psychological features of students, such as ARTHUR (Gilbert & Ham 1999) that considers three learning styles (visual-interactive, reading-listener, textual), CS388 (Carver et al. 1996) and MAS-PLANG (Pea et al. 2002) that use Felder and Silverman styles; and the INSPIRE system (Papanikolaou et al. 2003) that uses the styles proposed by Honey and Mumford (Honey & Mumford 1992). Different techniques have been used to represent student models such as rules (Jeremic & Devedzic 2004), BN and case-based reasoning (Pea et al. 2002). Particularly, in our research group, neural networks (Villaverde et al. 2006) have also been tested. Although the precision obtained at detecting learning styles with this technique was good, neural networks are not as easily understood as BN, as it is hard to explain and understand the values taken by nodes in a neural network and the structure of the network. The graphic form of BN makes them more comprehensible and intuitive. As regards BN, ANDES (Gertner & VanLehn 2000) and SE-Coach (Conati et al. 2002) use this technique to model students knowledge in Physics. IDEAL (Shang et al. 2001) uses this technique to categorize students into novice, beginner, intermediate, advanced or expert. In the study by Arroyo and Woolf (2005), the authors build a Bayesian model to detect a students hidden attitudes, perception towards the e-learning system and amount learned from observable student behaviour recorded in a log le. In Gamboa and Fred (2001), the authors use BN to assess students state of knowledge and learning preferences in an intelligent tutoring system. Our work uses BN to model a students learning style, an aspect not considered in the previous Bayesian student models.
Conclusions and future work

Web-based courses. The proposed model has been evaluated during three semesters with good results. The information obtained from these experiments not only enabled us to adjust our initial model by incorporating new variables and by adjusting the probability values in the BN, but also to enhance the quality of the Web-based course improving in this way students learning. The precision of our technique to automatically detect students learning styles increased for all the dimensions modeled. We are currently working on an intelligent agent that will use the information about a students learning style to provide him/her personalized assistance with the Web-based course, with the aim of improving the students performance. We will consider both providing assistance to supplement the negative aspects of the students learning style (for example, by suggesting intuitive learners to revise carefully the exam before handing it out) and providing assistance without intending to change the learning style.
Supplementary material

The following supplementary material is available for this article: Table S1. Experimental results (First experiment). Table S2. Experimental results (Second experiment). Table S3. Experimental results (Third experiment). This material is available as part of the online article from: DOI: 10.1111/j.1365-2729.2007.00262.x Please note: Blackwell Publishing are not responsible for the content or functionality of any supplementary materials supplied by the Authors. Any queries (other than missing material) should be directed to the Correspondence for this article.

Notes
1 2

http://www.engr.ncsu.edu/learningstyles/ilsweb.html Detailed information about how students used the system can be found in Experimental results section. 3 http://www.hugin.com

References

In this article we have presented an enhanced Bayesian model to infer the learning style of students that take

Allinson C. & Hayes J. (1996) The cognitive style index: a measure of intuition analysis for organizational research. Journal of Management Studies 33, 119135.

2007 The Authors. Journal compilation 2007 Blackwell Publishing Ltd

Students learning styles in Web-based courses

315

Arroyo I. & Woolf B. (2005) Inferring learning and attitudes from a Bayesian network of log le data. In AIED 05, 12th International Conference on Articial Intelligence in Education, 1822 July 2005, Amsterdam. Brusilovsky P. & Peylo C. (2003) Adaptive and intelligent web-based educational systems. International Journal of Articial Intelligence in Education 13, 159172. Carver C., Howard R. & Lavelle E. (1996) Enhancing student learning by incorporating learning styles into adaptive hypermedia. In ED-MEDIA96 World Conference on Educational Multimedia Hypermedia, pp. 118123. AACE, Charlottesville, VA. Chen S., Magoulas G. & Dimakopoulos D. (2005) A exible interface design for web directories to accommodate different cognitive styles. Journal of the American Society for Information Science and Technology 56, 7083. Cofeld F., Moseley D., Hall E. & Ecclestone K. (2004) Learning Styles and Pedagogy in Post-16 Learning: A Systematic and Critical Review. Learning and Skills Research Center, London. Conati C., Gertner A. & VanLehn K. (2002) Using Bayesian networks to manage uncertainty in student modeling. Journal of User-Modeling and User-Adapted Interaction 12, 371417. Dufresne A. & Turcotte S. (1997) Cognitive styles and its implications for navigation strategies. In Proceedings of the World Conference on Articial Intelligence and Education AI-ED97, 1822 August 1997, Kobe, Japan. Felder R. & Brent R. (2005) Understanding student differences. Journal of Engineering Education 94, 5772. Felder R. & Silverman L. (1988) Learning and teaching styles. Journal of Engineering Education 78, 674681. Ford N. & Chen S. (2000) Individual differences, hypermedia navigation, and learning: an empirical study. Journal of Educational Multimedia and Hypermedia 9, 281311. Gamboa H. & Fred A. (2001) Designing intelligent tutoring systems: a Bayesian approach. In ICEIS 01, 3rd International Conference on Enterprise Information Systems, 710 July 2001, Setbal, Portugal. Garcia P., Amandi A., Schiafno S. & Campo M. (2007) Evaluating Bayesian networks precision for detecting students learning styles. Computers and Education 49, 794 808. Gertner A. & VanLehn K. (2000) Andes: a coached problem solving environment for physics. In Intelligent Tutoring Systems, ITS 2000, 1923 June 2000, Montral, Canada, pp. 131142. Gilbert J. & Ham C. (1999) Adapting instruction in search of a signicant difference. Journal of Network and Computer applications 22, 149160.

Honey P. & Mumford A. (1992) The Manual of Learning Styles, 3rd edn. Peter Honey, Berkshire. Jensen F. (2001) Bayesian Networks and Decision Graphs. Springer Verlag, New York. Jeremic Z. & Devedzic V. (2004) Design patters ITS: student model implementation. In IEEE International Conference on Advanced Learning Technologies, 30 August1 September 2004, Joensuu, Finland. Kolb D. (1984) Experiential Learning: Experience as the Source of Learning and Development. Prentice Hall, New York. Litzinger M. & Osif B. (1993) Accommodating diverse learning styles: designing instruction for electronic information sources. In What is Good Instruction Nows? Library Instruction for the 90s (ed. L. Shirato). Pierian Press, Ann Arbor, MI. Olesen K., Lauritzen S. & Jensen F. (1992) Hugin: a system creating adaptive causal probability networks. In 8th Annual Conference Uncertainty in Articial Intelligence, 1719 July 1992, Stanford, CA. Papanikolaou K., Grigoriadou M., Knornilakis H. & Magoulas G. (2003) Personalizing the interaction in a webbased educational hypermedia system: the case of INSPIRE. User Modeling and User-Adapted Interaction 13, 213267. Paredes P. & Rodriguez P. (2004) A mixed approach to modelling learning styles in adaptive educational hypermedia. In Proceedings of 3rd IASTED Conference on Web-Based Education (WBE2004), 1618 February 2004, Innsbruck, Austria. Pearl J. (1988) Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann, San Mateo, CA. Pea C., Marzo J., de la Rosa J. & Fabregat R. (2002) Un sistema de tutora inteligente adaptativo considerando estilos de aprendizaje. In IE 2002, 2022 November 2002, Vigo, Spain. Shang Y., Shi H. & Chen S. (2001) An intelligent distributed environment for active learning. ACM Journal on Educational Resources in Computing 1. Doi: 10.1145/384055. 384059 Villaverde J., Godoy D. & Amandi A. (2006) Learning styles recognition in e-learning environments with feed-forward neural networks. Journal of Computer Assisted Learning 22, 197206. Walters D., Egert C. & Cuddihy E. (2000) Learning styles and Web-based education: a quantitative approach. In Proceedings of 9th Annual FACT Conference on Instructional Technology, 30 May1 June 2000, Buffalo, NY, pp. 115177.

2007 The Authors. Journal compilation 2007 Blackwell Publishing Ltd