As discussed in Chapter 2, there are conflicting opinions about whether learning styles are fixed traits or whether they change over time, by subject or the
debate – whether the modelling of learning styles should be static (measured at the start of the learning) or dynamic (modelled periodically or continuously). There are two main methods for ITS to model learning styles: explicitly using the measuring instrument defined by the learning styles model, or implicitly using learner behaviour in the ITS.
3.1.1 Collaborative Modelling Using Questionnaires
The simplest way to measure learning styles is using the formal assessment
described in the adopted learning styles model, normally a questionnaire (Wang et al., 2006; Spallek, 2003). However completing questionnaires is onerous for students, who do not always lend enough time or attention to complete them accurately. It is also difficult to avoid unintended influences in the questions, with some answers perceived as being better (Popescu, 2009). Also, as the modelling is static, if learning styles change over time or subject, the student model will not be accurate. These problems can lead to an unreliable student model (Yannibelli et al., 2006). The „open model‟ approach allows students to modify their profile directly, and has been used to extend static modelling using questionnaires (Papanikolau et al., 2003). Whilst the open model approach gives increased learner control and feedback on the quality of the system model, it also increases the learner‟s workload and relies on the learner‟s understanding and knowledge of their preferred learning style.
Some examples of ITS that use questionnaires to model learning styles are: CS383 (Carver et al., 1999) was one of the first AHS to adapt to learning styles, modelling three dimensions of the Felder and Silverman model using the Index of Learning Styles (ILS) (Felder and Soloman, 1998).
CooTutor (Wang et al., 2006) models the Felder and Silverman learning styles using the ILS questionnaire.
INSPIRE (Papanikolau et al., 2003) uses the Learning Styles Questionnaire to model learning styles (Honey and Mumford, 1992). Learners can also directly modify the student model to reflect their preferences.
AES-CS (Triantafillou et al., 2004) models the field dependence or independence (Witkin, 1962) of learners using a questionnaire. Learners may also alter the behaviour of AES-CS by changing options such as the amount of feedback given. iWeaver (Wolf, 2003) uses a questionnaire to initialise the model of the Dunn and Dunn (1974) learning styles. Learners are given an explanation of their
learning style and recommendations for style of resources, but may adjust the model by choosing other styles of resource. After each module, learners give feedback on the learning resources they use, with a ranked list which is used to adjust the student model.
In summary, although collaborative modelling of learning styles using
questionnaires is the simplest method, the model will only be accurate if students lend enough time and attention to complete the questionnaire properly (Garcia et al., 2007).
3.1.2 Automatic Modelling Using Learner Behaviour
Implicit modelling of a student‟s learning style involves building and updating a student model automatically based on the student‟s behaviour and actions while they use an ITS for learning (Villaverde et al., 2006; Stash and De Brau, 2004; Garcia et al., 2007). Modelling learning style dynamically and continually updating the student model enables an ITS to adapt to changes in learning style over time or for different subjects. Implicit modelling removes any requirement for input by the student so they can concentrate on their learning task; however it is difficult to extract enough reliable information to build a robust student model. Some ITS have overcome this problem by designing interfaces with the goal of collecting data to model learning styles (e.g. Cha et al., 2006). However the main goal of an ITS is to intelligently help students to learn, so interface design should focus on promoting learning. Another way to overcome the problem of reliability is to adopt a mixed-modelling approach, initially modelling learning style using a questionnaire and then dynamically
updating the model (e.g. Paredes and Rodriguez, 2004).
The types of learner behaviour used to model learning style include navigation and browsing patterns, the choice of resources (including time spent and frequency of access), the use of chat forums and test performance (Popescu, 2009).
There have been many different approaches to the automatic modelling of learning styles, including:
Bayesian networks are probabilistic models that have been used to model the relationships between learning styles and behaviour factors. Garcia et al. (2007) used Bayesian networks to infer student learning styles from a history of their behaviour in using the ITS. Three dimensions of the Felder and Silverman (1988)
model were modelled, with precisions of 58-77%. Enhancing the Bayesian model improved precisions to 66-80% (Garcia et al., 2008). EDUCE (Kelly and
Tangney, 2004) offers different resources styled using four of the Gardner (1983) multiple intelligences, and uses Naïve Bayes to predict which resources students prefer, based on past choices.
Artificial neural networks are computational models inspired by the neural structure of the brain, which have been used to classify student learning styles based on behaviour. Villaverde et al. (2006) used a neural network to determine student learning style for three dimensions of the Felder and Silverman (1988) model. The neural network analyses recent student behaviour in an ITS to automatically model learning style, achieving an accuracy of 69.3%. Hsu et al. (2010) used fuzzy inference rules to construct a neural network that identifies the relationship between learning activities and learning style. However, neural networks are less reliable for large amounts of input data (i.e. behaviour factors), (e.g. only ten behaviour factors (input neurons) were used by Villaverde et al., 2006) so this may not be enough to accurately model learning style. The opaque nature of neural networks also means that no information is learned about which behaviour factors are most significant in predicting a learning style.
Genetic algorithms are adaptive heuristic search algorithms that mimic the process of evolution by natural selection. Yannibelli et al. (2006) adopted a genetic algorithm approach to model three dimensions of the Felder-Silverman (1988) learning styles model based on student behaviour in an ITS.
Rule based methods involve modelling learning styles using rules that map patterns of behaviour extracted from learning styles models to learner behaviour. The DeLes tool (Graf et al., 2009) uses a rule-based method to infer student learning styles from their behaviour in a general Learning Management System (e.g. Moodle, 2011). DeLes models students using the Felder-Silverman (1988) model, and achieved precisions of 73-79%. WELSA (Popescu, 2009) also uses rule-based modelling (based on over 100 patterns of behaviour gathered from choice of learning resources, navigation and communication) to model learning styles using the Unified Learning Style Model (Popescu et al., 2007).
Decision trees are models that can predict the value of a variable (e.g. learning style) based on a number of input variables (e.g. behaviour factors). Cha et al.
(2006) designed their ITS interface to capture behaviour related to the Felder and Silverman (1988) model, using decision trees and Hidden Markov Models to classify learning styles from learner choices and behaviour. Ozpolat and Akar (2009) used the NBTree algorithm to classify student learning styles from the content of learning objects rather than behaviour (using keywords matched in Internet search terms). Chen and Liu (2008) used decision trees and K-means clustering to automatically identify cognitive styles from learning patterns. Other research has involved plotting clusters of types of learners against
behaviour factors, including a „dead band‟ where the learning style is classified as „unknown‟ (as in Sanders and Bergasa-Suso, 2010). Klasnja-Milicevic et al. (2011) clustered learners based on their learning style and then used the
AprioriAll pattern mining algorithm to extract behavioural patterns from log files. By comparing learner behaviours to each cluster, learning style was identified and learning material recommended.
Central to all of these automatic modelling methods is the capture of behaviour characteristics during use of the ITS. It is difficult to decide on the most appropriate behaviour to model and selecting typical behaviours that discriminate between learning styles requires a detailed analysis of the chosen learning styles model. Even then, students do not always behave stereotypically as suggested by learning styles models (Coffield et al., 2004a; Garcia et al., 2007). There are clear differences in the number of behaviour characteristics used by ITS to model learning styles – e.g. Garcia et al. (2007) capture 11, Cha et al. (2006) capture 58 and Popescu (2010) captures over 100. This does not always lead to different levels of precision in modelling learning styles, as different modelling methods and learning styles models have different requirements.
This section has described the main methods of modelling learning styles, either from explicit learner information gathered from learning styles questionnaires or implicitly from learners‟ behaviour while using an ITS. The next section outlines how adaptation to individual learning styles is approached.