OBOKO Non-obtrusive determination of learning styles in adaptive web-based learning

1.0 Abstract The Internet has revolutionalized the lives of people around the world. A number of services have become available courtesy of the Internet. One of these services is webbased learning, which may be used to educate students in schools and colleges as well as for staff training in organizations and for lifelong learning - anybody can participate in web-based learning. Hence, the people who use this service are diverse and therefore there is need to add an adaptive dimension to educational hypermedia, in efforts to address individual differences. Adaptivity is made with regard to different learner traits such as cognitive styles, learning styles, level of knowledge and attitude. Picking on learning styles, there are many learning style models existing. One of them is the VARK model, which encourages the provision of appropriate multi-media content to the learner. The learning style of a learner can vary over time as learning with hypermedia takes place. A machine-learning algorithm can be used to predict the learners learning style at any point in time during learning. This makes it possible to provide learning depending on a learners style, hence improving various aspects of learning. 2.0 Introduction This paper reports work on the development of a prototype of a web-based e-learning system for teaching c++ programming to first year Master of Science (MSc) in Information Systems at the School of Computing, University of Nairobi. A course was developed for the purpose. The course was developed to take care of different types of learners, based on their learning styles, with components developed to suit those who are inclined towards using visual, text, audio and activity based material. The sequencing of the materials was guided by the Elaboration Theory (Raigeluth, 1999) and the development of specific components such as activities, theory presentations and instances of the concepts for the different levels of performance i.e. remember, use and find, was based on the Component Display Theory (CDT) by David Merill (1983; 1987 in De Villiers, 2002). A key aspect of the study was the demonstration of the capturing of the dynamic learning style of the learner. 3.0 Review of related literature There has been a change of focus from web-based learning systems that just turn pages of content to systems which present learning material in such a way as to make sure that the learning needs of the learners are actually satisfied (Brusilovsky, 1999; Papanikolaou et al, 2002). This is especially important in modern learning which puts emphasis on the learners individuality in terms of the learners previous knowledge, views of the world, and preferred way of processing information and experiences from learning. The shift has led to the need for multiple kinds of learning, which require multiple methods of instruction. This should enable learners, in their uniqueness, to construct their own knowledge as opposed to the information merely being conveyed to the learners 1

(Papanikolaou et al, 2002). The uniqueness of the learners is met through making the web-based learning systems adaptive (Brusilovsky, 1999). In adaptive learning, the teaching strategy is adapted based on the contents of a learners model, which are generated based on the learners actions (e.g. visited pages) and on the success of the strategy the learner is currently employing determined by the learners performance on the tests he/she submits. The contents of a learner model may include the learners knowledge level and individual traits such as learning styles. The learning system generates and provides individualized web content by exploiting this learner model. Individuals perceive and process information in very different ways (Papanikolaou et al, 2002). 3.1 Learner traits Adaptivity may be based on different learner traits. One approach is to consider the learners knowledge level. The learning system restricts the domain knowledge at the beginning of the interaction, when the learner is assumed to be a novice (Brainsford, et al., 1999) and enriching it progressively, following improvement in their performance. The content presented to the learner may also be varied through adaptive navigation support, e.g. through adaptive link annotation or link hiding, with the learner being allowed to any part of the course, should need arise (learners still retain control). Another learner trait to consider during adaptivity is the learners learning style. Learning styles deal with how individual learners prefer to learn. It indicates how a learner prefers to deal with information and experiences from learning. This preference applies to different content areas. There are many ways of categorizing learning styles. One of them is based on the learners instructional and environmental preference (Claxton and Murrell, 1987). With this approach, learning styles are classified based on preferred sensory perception into auditory, aural, visual, tactile, etc. One model fitting with this approach is the VARK (Visual, Aural, Read/Write and Kinesthetic) Model. Each of the four learning styles corresponds to the type of media preferred by a learner. Learners with the visual style prefer information to arrive in the form of graphs, charts, various diagrams, etc. They are also sensitive to the use of colours. For those with the aural style, speech is the most effective way of receiving information. Learners with this preference learn best from lectures, discussions, etc. For Learners with the read/write learning style, the preferred form of received information is written or printed words. Learners with this preference learn best from textbooks, lecture notes, handouts, etc. Learners with the kinesthetic learning style prefer concrete, multi-sensory experience. They learn by doing. Learners with this preference learn best from practical sessions, field trips, experiments, role playing or simulation i.e. real life experiences. 3.2 Machine learning algorithms Machine learning refers to the use of various algorithms to enable a machine (read computer system) to learn from examples provided to it during training (Mitchell, 1997). The goal is for the machine to be able to carry out certain tasks (e.g. classification) and get better at performing these tasks with more training (i.e. the machine learns).

Examples of these algorithms include Neural Networks, Nave Bayes Classifier, Instance Based learners and Reinforcement learners. Machine learning algorithms may be used to predict a learners knowledge level or learning style, based on previously recorded actions of learners in the course of learning, such as links followed, tests submitted, errors made, etc. For example, a Nave Bayes Classifier is used to predict a learners learning style using the Felder & Silverman Index of Learning Style (Kelly & Tangey, 2005), Fuzzy Logic is used to predict a learners learning style and knowledge level in Hong, H. & Kinshuk (2004) and fuzzy logic is used to predict a learners knowledge level (Papanikolaou, et al, 2002).

3.3 The Nave Bayes Classifier Nave Bayes Classifiers have proved to be powerful tools for solving classification problems (Mitchell, 1997). Such problems involve fitting a new i.e. previously unseen case into one of the existing categories. These classifiers have been successfully applied in such areas as text and document classification, medical diagnosis, and electronic performance support systems. A Nave Bayes Classifier is composed of a single class variable, whose values are the possible outcomes or classes for the specific example problem under study, and a set of feature variables, modeling the features that provide for distinction among various classes. The feature variables are assumed to be mutually independent, given the class variable, i.e. P(X/C) = i=1 n P(Xi/C) Where X = (X1, Xn) is a feature vector and C is a class. Nave Bayes classifiers are typically learned from data (training examples). Learning such a classifier amounts to establishing the prior probabilities of the different classes and estimating the conditional probabilities of the various features given each of the classes. These classifiers assign the most likely class to a given training example, as described by its feature variables. For a learner classification problem, this might be classes of knowledge level such as beginner, intermediate and advanced (Papanikolaou et al, 2002; Tsiriga & Virvou, 2004) or aural, visual, read/write and kinesthetic learning styles (Wolf, 2004) or activist, pragmatist, theorist and reflector learning styles (Papanikolaou et al, 2002) or one of the many other ways of classifying learners. 3.4 Adaptivity Adaptivity refers to the process by which a system adapts its output using some data or knowledge about the learner such as the learners knowledge level or learning styles (Brusilovsky, 1999). Examples of technologies for adaptivity are curriculum sequencing, adaptive navigation and adaptive presentation.

Curriculum sequencing The goal of the curriculum sequencing technology (also referred to as instructional planning technology) is to provide the learner with the most suitable individually planned sequence of knowledge units/concepts to learn and sequence of learning tasks (examples; questions and, problems) to work with (Brusilovsky, 2003). Adaptive navigation - The goal of the adaptive navigation support technology is to support the learner in hyperspace orientation and navigation by changing the appearance of visible links. In particular, the system can adaptively sort, annotate, or partly hide the links of the current page to make easier the choice of the next link to proceed to (Brusilovsky, 2003). Adaptive presentation - The goal of the adaptive presentation technology is to adapt the content of a hypermedia page to the user's goals, knowledge and other information stored in the user model. In a system with adaptive presentation, the pages are not static, but adaptively generated or assembled from pieces for each user. Adaptive presentation is very important in WWW context where the same "page" has to suit to very different learners (Brusilovsky, 2003). To realize adaptivity based on learning styles, one of the methods is the application of multiple teaching strategies to suit different learning styles (Papanikolaou, et al., 2002). For the VARK learning styles model, learning may be offered using text, sound, learner activities and tasks, or through multiple media, depending on the learners learning style (Wolf, 2004). 3.5 Course Design In designing a course, different instructional design theories can be considered for providing guidelines. Examples of these are Elaboration Theory (ET) by Raigeluth (1999) and Component Display Theory (CDT) by Merill (1983; 1987 in De Villiers, 2002). ET is used to guide the organization of the main concepts in terms of the selection and sequencing of concepts. CDT is used to guide on specific strategies for guiding instruction of a single topic/concept such as definitions, examples and practice (Merill 1983; 1987 in De Villiers, 2002). In order to make page generation possible, learning objects are needed connectionist approach. A course is broken down into components, which can be connected together to form pages suited to a given learners learning style (Papanikolaou, et al, 2000). These components have attributes such as the main discipline e.g. object oriented programming, concept it teaches, sub- concepts, end-user type( e.g. aural), document format, difficulty level(e.g. novice), interactivity level, didactic context e.g. university level, course level e.g. introductory. These attributes are used to determine when and for which learner a piece of course content is to be presented. 4.0 Key features of the web-based learning system 4.1 Design of the course

ET was used to guide the selection of OOP concepts presented to the learners. It also guided the sequencing of concepts i.e. from simple to complex concepts, with the details coming later. The theory also guided the design of the course in terms of showing which OOP concepts are pre-requisites to others. CDT was used to guide the development of definitions, examples and practice for every concept. Practice included a short test at the end of the concept as well as programming tasks. The basic sequence of definitions, examples and practice was followed for every concept. Components were developed in such a way that all the end-user types with respect to the VARK model could be provided with suitable presentations: a) Aural Presentation The aural presentation was created using a combination of RealText files and audio files which were audio recordings of the learning material placed on the web-based learning system. b) Visual Presentation Pictures and diagrams were created to represent the content. Code segments provided and the visual learners were required to complete the code by combining different pieces into a whole. c) Read/Write Presentation For this preference, full text representation of the learning material was used. Colour coding and formatting was also employed, for example, showing important points using different colours and formatting. d) Kinesthetic Presentation For this preference, the learners were given the opportunity to write and compile c++ code online. An open source compiler was installed on the web server. 4.2 Learner modeling Nave Bayes Classifier was used to predict the current learning style orientation of the learner. The visited links of the learner, whether a given type of learning material was reused, the attempted questions, and the number of attempts from the learner per question was fed to the algorithm to predict the current learning style. It was assumed that a learners learning style can vary over the course of learning. Therefore, a learners learning style was not determined and fixed at the beginning by having learners fill a long questionnaire. Learners often employ different strategies depending on the tasks they are facing (Wolf, 2004). 4.3 Adaptivity Adaptivity in the learning system was realized using adaptive presentation. The appropriate type of learning content was generated for the learner, depending on the learners predicted learning style. A database query would receive the learning style,

knowledge level (beginner, refresher, and reference), the current concept, sub-topic of the concept currently being studied by the learner and the current state (definition, example, and exercise). Appropriate content is retrieved and passed to the display function. The system also has aspects of navigation support since a learner is only allowed to see previously studied concepts. Tracking of user activity was important. The following information was collected as a learner interacted with the system: whether a given resource was the first one to be chosen at the start of learning the concept, was the resource reused in the course of the learning session, whether or not a question was attempted after using a particular resource, whether or not the learner answered test questions satisfactorily after having learned using a certain resource, and whether or not a certain resource was the only one used throughout the learning session. The Nave Bayes Classifier used this information to predict the resource the user would use next. This classifier, unlike others like Neural Networks, requires very little data to begin its predictions. The predicted resource was validated against the resource the user actually chose. The learners were allowed to have control over what they learned (De Villiers, 2002; Papanikolaou et al, 2001). Links were located at the bottom of each page throughout the course and labeled visual, aural, read and kinesthetic (according to VARK model). These links were used to override the predicted learning style. Then, the display function was activated, and content was displayed according to the selected learning style, concept and sub-concept currently being viewed and the state. To aid the learners in making the choices, a full explanation of the different types of learning material was given at the beginning of the course i.e. visual for pictorial presentation of content, aural for audio presentation of content, read for full text presentation of content and kinesthetic for a do it yourself oriented approach to learning. 4.4 System Architecture Learning content is presented to users in the form of web-pages. As the learner interacts with the web-pages, information about the learner is collected and stored in the database. The classifier attempts to predict which resource the user would prefer to view on subsequent pages. The display function presents various multimedia representations to the learner, based on the predicted learning style or the learning style expressly specified by the learner.

Display Function

Web Pages Presented to the user

User

Bayesian Predictions passed to display function Bayesian Classifier Display Parameters Reads usage data Usage information Stores in a database User Data

Represents the virtual links within the system

Figure 1: System Architecture 5.0 Design of the Study The entire class of first year MSc Information systems, in May, 2006 was targeted. It had 40 students. The students participated in the study on a voluntary basis. The aim was to observe the accuracy of predicting a learners learning style (and hence the resource the learner was likely to use next). 5.1 Results and Discussion The data on the systems predictions were collected in a table with this format: Username Predicted_class Chosen_resource Correct Concept state Track ambalu V R 0 1 3 1 ambalu V V 1 1 4 1 ambalu V V 1 1 5 1 Table 1: Resource predictions for students

Each row shows per-page predictions of learning styles/resources made by the classifier. If the predicted and actual resources are the same, a value of 1 was entered in the column Correct. Concept, State and Track described the concept, particular part of concept being followed and the learners level of mastery of the course (beginner, refresher or reference). For all the learners who participated, the sum of correct out of all the entries was made. The accuracy of the classifiers predictions was 83%. 5.2 Conclusion From the study, it was possible to confirm that users are different and require different kinds of learning resources in the learning environment. The learners preference of resources for learning can change in the course of learning. Therefore, fixing it at the start of learning might adversely affect learners. It is possible to predict a learners learning style and therefore, the preferred learning resources, non-obtrusively, using machine learning techniques. This spares learners the experience of filling in long questionnaires at the beginning and also being fixed in one learning style throughout the course of learning (since filling the questionnaire time and again may be impractical), yet the learning style can change. This has been achieved in this study, as demonstrated by the high level of accuracy of the predictions 5.3 Future Work The study can be extended to cover the automatic classification of the learning style classifications. This is to make it possible to draw authoritative generalizations about the effectiveness of non-obtrusive determination of learning styles. Testing should also be extended to cover other content areas, as well as students at different stages of learning and backgrounds besides object oriented programming.

References Brusilovsky, P. (1999) Adaptive and Intelligent Technologies for Web-based Education. In C. Rollinger and C. Peylo (eds.), Special Issue on Intelligent Systems and Teleteaching, Knstliche Intelligenz journal, 4, 19-25 Brusilovsky, P. (2003) Developing adaptive educational hypermedia systems: From design models to authoring tools. In: T. Murray, S. Blessing and S. Ainsworth (eds.): Authoring Tools for Advanced Technology Learning Environment. Dordrecht: Kluwer Academic Publishers, 377-409 Claxton, Charles S and Murrell Patricia H. (1987). Learning Styles : Implications for Improving Educational Practices. Association for the study for Higher Education De Villiers, M (2002). The Dynamics in the Theory and Practice of Instructional Design PhD Thesis at University of Pretoria Hoffman, S. (1997). Elaboration Theory and Hypermedia: Is there a Link? Educational Technology, (37)1, 57-64. http://www.britishcouncil.org/ecs/events/2005/0203/details/market.htm#contents Kelly D & Tangney, B (2005). First Aid For You. Proceedings of the Seventh International Conference on ITSs Mitchell, T (1997). Machine Learning. McGraw-Hill, New York. Papanikolaou, K.A., Grigoriadou, M, Kornilakis, H. and Magoulas, G.D. (2002). Personalizing the Interaction in a Web-based Educational Hypermedia System: The case of INSPIRE. Reigeluth, C. M. (1999). (Ed.),A new paradigm of instructional theory, Volume II, pp. 529). Mahwah, NJ: Lawrence Erlbaum Associates. Savery, R & Duffy, T(1995). Problem Based Learning: An Instructional model and its constructivist framework. Educational Technology, 35, 1995 pp 31-38 Stoyanov, S (2001). Mapping in Educational and Training Design. Doctoral Dissertation, University of Twente Tsiriga, V & Virvou, M (2004). A Framework for the Initialization of Student Models in Web-based Intelligent Tutoring Systems. User Modelling and User-Adapted Interaction 14 pp 289-316 Wolf, C (2003). iWeaver: Towards 'Learning Style'-based e-Learning in Computer Science Education. Faculty of Education, Language and Community Studies RMIT University, Australia