# Introduction

eople have different views upon the same situation, the way they perceive and estimate the world is different. So their responses to around environment are also different. For example, look at the way students prefers to study a lesson. Some have a preference for listening to instructional content (socalled auditory learner), some for perceiving materials as picture (visual learner), some for interacting physically with learning material (tactile kinesthetic learner), some for making connections to personal and to past learning experiences (internal kinesthetic learner). Such characteristics about user cognition are called learning styles but learning styles are wider than what we think about them.

Learning styles are defined as the composite of characteristic cognitive, affective and psychological factors that serve as relatively stable indicators of how a learner perceives, interacts with and responds to the learning environment. Learning style is the important factor in adaptive learning, which is the navigator helping teacher/computer to deliver the best instructions to students.

There are many researches and descriptions about learning style but only minorities of them are valuable and applied widely in adaptive learning. The descriptions of learning style (so-called learning style models) are categorized following criteria: model)

In section 2, we discuses about such learning style families. In general, learning styles are analyzed comprehensively in theory of psychology but there are few of researches on structuring learning styles by mathematical tools to predict/infer users' styles. Former researches often give users questionnaires and then analyze their answers in order to discover their styles but there are so many drawbacks of question-and-answer techniques, i.e., not questions enough, confusing questions, users' wrong answers? that such technique is not a possible solution. It is essential to use another technique that provides more powerful inference mechanism. So, we propose the new approach which uses hidden Markov model to discover and represent users' learning styles in section 4, 5. We should pay attention to some issues of providing adaptation of learning materials to learning styles concerned in section 3. -Constitutionally based learning styles and preferences (Dunn and Dunn) -The cognitive structure (Witkin, Riding) -Stable personality type (Myers-Briggs)


# II.


# Learning Style Families

-Flexibly learning preferences (Kolb, Honey-Mumford, Felder-Silverman, Pask and Vermunt model) Environmental: incorporates user preferences for sound, light, temperature? Emotional: considers user motivation, persistence, responsibility? Sociological: discovers user preference for learning alone, in pairs, as member of group


# b) The Cognitive Structure

In this family, learning styles are considered as structural properties of cognitive system itself. So styles are linked to particular personality features, which implicates that cognitive styles are deeply embedded in personality structure. There are two models in this family: Witkin model and Riding model.

i


# . Witkin Model

The main aspect in Witkin model [Witkin, Moore, Goodenough, Cox 1997] is the bipolar dimensions of field-dependence/field-independence (FD/FI) in which:

Field-dependence (FD) person process information globally and attend to the most salient cues regardless of their relevance. In general, they see the global picture, ignore details and approach the task more holistically. They often get confused with non-linear learning, so, the require guided navigation in hypermedia space. Field-independency (FI) person are highly analytic, care more inherent cues in the field and are able to extract the relevant cues necessary to complete a task. In general, they focus on details and learn more sequentially. They can set learning path themselves and have no need of guidance.


# ii. Riding Model

Riding model [Riding, Rayner 1998] identifies learning styles into two dimensions: Wholist-Analytic and Verbalizer-Imager.

Wholist-Analytic dimension expresses how an individual cognitively organize information either into whole or parts. Wholist tends to perceive globally before focusing on details. Otherwise, analytic tends to perceive everything as the collection of parts and focusing on such parts.

Verbalizer-Imager dimension expresses how an individual tends to perceive information, either as text or picture. Verbalizer prefers to text. Imager prefers to picture.


# c) Stable Personal Type

The models in this family have a common focus upon learning style as one part of the observable expression of a relatively stable personality type. We will glance the famous model in this family: Myers-Briggs Type Indicator. Based on four stages, there are four learning styles: accommodating, assimilating, diverging and converging. Each couple of these stages constitutes a style, for example, CE and AE combine together in order to generate accommodating style. conceptualization and reflective observation. Learners respond to information presented in an organized, logical fashion and benefit if they have time for reflection. A typical question for this style is "What?" conceptualization and active experimentation. Learners respond to having opportunities to work actively on well-defined tasks and to learn by trialand-error in an environment that allows them to fail safely. A typical question for this style is "How?" experience and reflective observation. Learners respond well to explanations of how course material relates to their experience, their interests, and their future careers. A typical question for this style is "Why?"

ii.


# Honey and Mumford Model

According to Peter Honey and Alan Mumford [Honey, Mumford 1992]  The adaptive strategy (for learning style) is the sequence of adaptive rules which define how adaptation to learning styles is performed. Learning style strategies is classified into three following forms: materials) is presented in various types such as: text, audio, video, graph, picture? Depending on user's learning styles, an appropriate type will be chosen to provide to user. For example, verbalizers are recommended text and imagers are suggested pictures, graphs. This form support adaptation techniques such as: adaptive presentation, altering fragments, stretch text? navigation paths: The order in which learning materials are suggested to users is tuned with learning styles. For active learners, learning materials are presented in the order: activity?example?theory?exercise. For reflective learner, this order is changed such as: example?theory?exercise?activity. This form is corresponding to link adaptation techniques: direct guidance, link sorting, link hiding, link annotation.

Different learning tools are supported to learners according to their learning styles. For example, in Witkin model, FD learners are provided tools such as: concept map, graphic path indicator. Otherwise FI learners are provided with a control option showing a menu from which they can choose in any order (because they have high self-control).

There are two type of strategy: adaptive rules and is in three above forms.

to observe user actions and infer their learning styles. Thus, meta-strategy is applied in order to define strategy.

Our approach is an instructional meta-strategy that apply Markov model to infer users' learning styles. Before discussing about main techniques, it is necessary to glance over hidden Markov model.


# IV.


# Hidden Markov Model

There are many real-world phenomena (socalled states) that we would like to model in order to explain our observations. Often, given sequence of observations symbols, there is demand of discovering real states. For example, there are some states of weather: sunny, cloudy, rainy. Based on observations such as: wind speed, atmospheric pressure, humidity, temperature?, it is possible to forecast the weather by using Hidden Markov Model (HMM). Before discussing about HMM, we should glance over the definition of Markov Model (MM). First, MM is the statistical model which is used to model the stochastic process. MM is defined as below: cardinality is n. Let ? be the initial state distribution where ? i ? ? represents the probability that the stochastic process begins in state s i . In other words ? i is the initial probability of state s i , where 1 = ? ?S s i i ? one state from S at all times. The process is denoted as a finite vector P=(x 1 , x 2 ,?, x u ) whose element x i is a state ranging in space S. Note that x i ? S is one of states in the finite set S, x i is identical to s i . Moreover, the process must meet fully the Markov property, namely, given the current state x k of process P, the conditional probability of next state x k+1 is only relevant to current state x k , not relevant any past state (x k-1 , x k-2 , x k-3 ,?). In other words, Pr (x k | x 0 , x 1 ,?, x k-1 ) = Pr(x k | x k-1 ). Such process is called first-order Markov process. state based upon the transition probability distribution a ij which depends only on the previous state. So a ij is the probability that, the process change the current state s i to next state s j . The probability of transitioning from any given state to some next state is 1:
1 , = ? ? ? ?S s ij i j a S s
. All transition probabilities a ij (s) constitute the transition probability matrix A.

Briefly, MM is the triple ? S, A, ? ?. In typical MM, states are observed directly by users and transition probability matrix is the unique parameters. Otherwise, Hidden Markov Model (HMM) is similar to MM except that the underlying states become hidden from observer, they are hidden parameters. HMM adds more output parameters which are called observations. Each state (hidden parameter) has the conditional probability distribution upon such observations. HMM is responsible for discovering hidden parameters (states) from output parameters (observations), given the stochastic process. The HMM have further properties as below:

produces observations correlating hidden states. Suppose there is a finite set of possible observations ?"¨ = {? 1 , ? 2 ,?, ? m } whose cardinality is m.

given observation in each state. Let b i (k) be the probability of observation ? k when the second stochastic process is in state s i . The sum of probabilities of all observations which observed in a certain state is 1, -Instructional strategy is itself, which contains 
1 ) ( , = ? ? ? ?? ? k k b S i i . All

# Uncovering problem and Viterbi algorithm

Given HMM ? and a sequence of observations O = {o 1 ? o 2 ??? o k } where o i ? ?"¨ , how to find the sequence of states U = {u 1 ? u 2 ??? u k } where u ? S so that U is most likely to have produced the observation sequence O . This is the uncovering problem: which sequence of state transitions is most likely to have led to this sequence of observations. It means to maximize the selection of U:
)] | [Pr( max arg ? O U .
We can apply brute-force strategy: "go through all possible such O and pick the one with the maximum" but this strategy is infeasible given a very large numbers f states. In this situation, Viterbi algorithm [Dugad, Desai 1996] is the effective solution. Instead of describing details of Viterbi algorithm, we only use it to predict learner's styles given observations about her/him.


# V. Applying hidden markov Model Into Modeling and Inferring Users' Learning Styles

For modeling learning style (LS) using HMM we should determine states, observations and the relationship between states and observations in context of learning style. In other words, we must define five components S, ?"¨, A, B, ?. Each learning style is now considered as a state. The essence of state transition in HMM is the change of user's learning style, thus, it is necessary to recognize the learning styles which are most suitable to user. After monitoring users' learning process, we collect observations about them and then discover their styles by using inference mechanism in HMM, namely Viterbi algorithm. Suppose we choose Honey-Mumford model and Felder-Silverman model as principal models which are presented by HMM. We have three dimensions: Verbal/Visual, Activist/ Reflector, Theorist/ Pragmatist which are modeled as three HMM(s): ? 1 , ? 2 , ? 3 respectively. For example, in ? 1 , there are two states: Verbal and Visual; so S 1 ={verbal, visual}. We have:
- ? 1 = ? S 1 , ?"¨ 1 , A 1 , B 1 , ? 1 ?. - ? 2 = ? S 2 , ?"¨ 2 , A 2 , B 2 , ? 2 ?. - ? 3 = ? S 3 , ?"¨ 3 , A 3 , B 3 , ? 3 ?.
We are responsible for defining states (S i ), initial state distributions (? i ), transition probability matrices (A i ), observations (?"¨ i ), observation probability matrices (B i ) through five steps 1. Defining states: each state is corresponding to a leaning style. S 1 = {verbal, visual}, S 2 = {activist, reflector}, S 3 = {theorist, pragmatist}. 2. Defining initial state distributions: we use uniform probability distribution for each ? i . ? 1 = {0.5, 0.5}; it means that Pr (verbal) = Pr (visual) = 0.5 ? 2 = {0.5, 0.5}; Pr(activist) = Pr(reflector) = 0.5 ? 3 = {0.5, 0.5}; Pr (theorist) = Pr (pragmatist) = 0.5 3. Defining transition probability matrices: we suppose that learners tend to keep their styles; so the conditional probability of a current state on previous state is high if both current state and previous state have the same value and otherwise. For example, 4. Defining observations. There is a relationship between learning object learned by users and their learning styles. We assign three attributes to each learning object (such as lecture, example?): ? Format attribute indicating the format of learning object has three values: text, picture, video . ? Type attribute telling the type of learning object has four values: theory, example, exercise, and puzzle . ? Interactive attribute indicates the "interactive" level of learning object. The more interactive learning object is, the more learners interact together in their learning path. This attribute has three values corresponding to three levels: low, medium, high . Whenever a student selects a learning object (LO), it raises observations depending on the attributes of learning object. We must account for the values of the attributes selected. For example, if a student selects a LO which has format attribute being text, type attribute being theory, activity attribute being low, there are considerable observations: text, theory, low (interaction). So, it is possible to infer that she/he is a theorist.
Pr(s i =verbal | s i-1 =verbal) = 0.7 is obviously higher than Pr(s i =verbal | s i-1 =verbal) = 0.
The dimension Verbal/Visual is involved in format attribute. The dimensions Activist/ Reflector and Theorist/ Pragmatist relate to both type attribute and interactive attribute. So we have: Theory, example, exercise, puzzle, low (interaction), medium (interaction), high (interaction) }
? ?"¨ 1 = { Text, picture, video } ? ?"¨ 2 = {
? ?"¨ 3 = { Theory, example, exercise, puzzle, low (interaction), medium (interaction) high (interaction) } 5. Defining observation probability matrices. Different observations (attributes of LO) effect on states (learning styles) in different degrees. Because the "weights" of observation vary according to states, there is a question: "How to specify weights?" If we can specify these "weights", it is easy to determine observation probability matrices.

In the Honey-Mumford model and Felder-Silverman model, verbal students prefer to text material and visual students prefer to pictorial materials. The weights of observations: text, picture, video on state Verbal are in descending order. Otherwise, the weights of observations: text, picture, video on state Visual are in ascending order. Such weights themselves are observation probabilities. We can define these weights as below:
? Pr(text | verbal) = 0.6, Pr(picture | verbal) = 0.3, Pr(video | verbal) = 0.1 ? Pr(text | visual) = 0.2, Pr(picture | visual) = 0.4, Pr(video | visual) = 0.4
There are some differences in specifying observation probabilities of dimensions Activist/Reflector and Theorist/ Pragmatist. As discussed, active learners are provided activity-oriented approach: showing content of activity (such as puzzle, game?) and links to example, theory and exercise. Reflective learners are provided example-oriented approach: showing content of example and links to theory, exercise and activity (such as puzzle, game?). The weights of observations: puzzle, example, theory, exercise on state Activist are in descending order. The weights of observations: example, theory, exercise, puzzle on state Reflector are in descending order. However, activists tend to learn high interaction materials and reflectors prefer to low interaction materials. So the weight of observations: low (interaction), medium (interaction), high (interaction) on state Activist get values: 0, 0, 1 respectively. Otherwise, the weight of observations: low (interaction), medium (interaction), high (interaction) on state Reflector get values: 1, 0, 0 respectively. We have: showing content of theory and links to example, exercise and puzzle; pragmatists are provided exercise-oriented approach: showing content of exercise and links to example, theory and puzzle. Thus, the conditional probabilities of observations: example, theory, exercise, puzzle, low (interaction), medium (interaction), high (interaction) on states: theorists, pragmatists are specified by the same technique discussed above.  
? Pr(puzzle | activist) = 0.4, Pr(example | activist) = 0.3, Pr(theory | activist) = 0.2, Pr(exercise | activist) = 0.1 Pr(low | activist) = 0, Pr(medium | activist) = 0, Pr(high | activist) = 1. ? Pr(example | reflector) = 0.4, Pr(theory | reflector) = 0.3, Pr(exercise | reflector) = 0.2, Pr(puzzle | reflector) = 0.1 Pr(low | reflector) = 1, Pr(medium | reflector) = 0, Pr(high | reflector) = 0.

# IV Version I

Now three HMM (s): ? 1 , ? 2 , ? 3 corresponding to three dimensions of learning styles: Verbal/Visual, Activist/Reflector, Pragmatist/Theorist are represented respectively in figure 2.


# An example for inferring student's learning styles

Suppose the learning objects that a student selects in session 1, 2 and 3 are LO 1 , LO 2 and LO 3 respectively. 


# Format


# Conclusion

HMM and Viterbi algorithm provide the way to model and predict users' learning styles. We propose five steps to realize and apply HMM into two learning style models: Honey-Mumford and Felder-Silverman, in which styles are considered states and user's selected learning objects are tracked as observations. The sequence of observations becomes the input of Viterbi algorithm for inferring the real style of learner. It is possible to extend our approach into other learning style models such as: Witkin, Riding, Kolb? and there is no need to alter main techniques except that we should specify new states correlating with new learning styles and add more attributes to learning objects.
![Learning styles in this family are fixed and difficult to change. This family has the famous model "Dunn and Dunn model" developed by authors Rita Dunn and Kenneth Dunn [Dunn, Dunn 2003]. With Dunn Dunn model, learning style is divided into 5 major strands: Their theoretical importance -Their wide spread use -Their influence on other learning style models -Learning style models are organized within the families such as:](image-2.png "and-")



Pask model developed by Pask [Pask 1976]-Converging (AC/AE): relies primarily on abstractstates that there are two learning styles:Wholist: Learners understand problems by buildingup a global viewSerialist: Learners prefer to details of activities, factsand follow a step-by-step learning procedure.-Diverging(CE/RO):emphasizesconcretev. Vermunt ModelAccording to Vermunt [Vermunt 1996], theauthor of this model, there are four learning styles:Active (also impulsive, extravert) learners are provided activity-oriented approach: showing2013content of activity and links to example, theory and exercise. Reflective (also introvert) learners areYearprovided example-oriented approach: showing?content of example and links to theory, exercise and activity.3Concrete experience (CE) Abstract conceptualization (AC) Accommodating Diverging III. Providing Adaptation of Learning Reflective observation (RO) Active Materials to Learning Styles Learning styles are discovered and explored in psychological domain but how they are incorporated into adaptive systems? We must solve the problem of "matching" learning materials with users' learning styles. The teacher must recognize styles of students and then provide individually them teaching methods associated personal learning materials (lesson, exercise, test?). Such teaching method is called learning strategy or instructional strategy or adaptive strategy. Although there are many learning style models but they share some common features, such as: the modality visual (picture)/visual (text) in Dunn and Dunn model is similar experimentation (AE) Assimilating Converging to verbalizer /imager dimension in riding model and verbal-visual dimension in Felder-Silverman model. Strategies are supposed according to common features of model because it is too difficult to describe comprehensively all features of model. Features of all models (learning styles) can be categorized into three groups: perception and understanding which are enumerated together with adaptive strategies as below: Perception group: This group related learners' perception includes:Activist Theorist The theorist/pragmatist dimension of Honey and Pragmatist Reflector Mumford model. Theorists are provided theory-oriented approach: showing content of theory and links to example, exercise and activity. Pragmatists are provided exercise-oriented approach: showing content of exercise and links to example, theory and activity. The accommodating/assimilating dimension of Kolb model is similar to application-directed/ meaning-oriented dimension of Vermunt model. The adaptive strategy for accommodating style is to provide application-based information to learners. Other-wise, theory-based information for assimilating style. Understanding group: This group related to the way learners comprehend knowledge includes: ? ? ? The global/analytical modality in Dunn and Dunn model is similar to wholist-analytic dimension in riding model, global/sequential dimension in Felder-Silverman model, wholist-serialist dimension in Pask model. Global (also wholist) learners are provided breadth-first structure of learning material. Otherwise, analytical (also analytic, sequential, serialist) learners are recommended depth-first structure of learning materials. For the breadth-first structure, after a learner has already known all the topics at the same level, other descendant topics at lower level are recommended to her/him. For theScience ( D D D D ) G Volume XIII Issue IV Version I Human Social Journal ofThe visual(picture) / visual(text) modality in Dunn and -Accommodating (CE/AE): emphasizes concrete experience and active experimentation. Learners Dunn model is similar to the verbalizer/imager dimension in Riding model and verbal-visualdepth-first structure, after a learner has already known a given topic T 1 and all its children (topic) at -Verbal/Visual. Verbal students like learning materials lower level, the sibling topic of T 1 (namely T 2 , atGlobalprefer to apply learning material in new situations so dimension in Felder-Silverman model. Instructionalsame level with T 1 ) will be recommended to her/him.that they solve real problems. A typical question for this style is "What if?" strategy is that the teacher should recommend textual materials to verbalizer and pictorial materialsThe FD/FI dimension in Wikin model is correlated -Sequential/Global. Sequential students structure with undirected/reproduction-oriented dimension in-Assimilating to imager.(AC/RO):prefersabstractVermunt model. FD learners are provided breadth-first structure of materials, guided navigation,illustration of ideas with visual materials, advanceorganizer and system control. FI learners areprovided depth-first structure of materials ornavigational freedom, user control and individualenvironment.Reflector: learners prefer to think about new information first before acting on it. Theorist: learners think things through in logical steps, assimilate different facts into coherent theory. Pragmatist: learners have practical mind, prefer to try and test techniques relevant to problems.iii. Felder-Silverman ModelThis model developed by Felder and Silverman[Felder, Silverman 1988] involves following dimensions: information only if they discussed it, applied it. A © 2013 Global Journals Inc. (US) -Active/Reflective. Active students understand -Sensing/Intuitive. Sensing students learn from The sensing/intuitive dimension in Felder-Silverman model is identical to the sensor/intuitive dimension in Myer Briggs Type Indicator. Sensing learners are recommended examples before expositions, otherwise, expositions before examples for intuitive learners.
probabilities of observations b i (k) constitute theobservation probability matrix B.S = {sun, cloud, rain}, ?"¨ = {dry, dryish, damp,soggy}weather todaysuncloudrainsun0.50.250.25weather yesterdaycloud0.40.20.4rain0.10.70.2Transition probability matrix Ahumiditydrydryishdampsoggysun0.60.20.150.05weathercloud0.250.250.250.25rain0.050.10.350.5Observation probability matrix BFigure 1 : HMM of weather forecast (hiddenstates are shaded)-There is the second stochastic process which-There is a probability distribution of producing a
1verbal visualerbal 0.70.3visual 0.30.7
2D D D D ) GScience (Human SocialJournal ofGlobalText Picture VideoVerbal 0.60.30.1Visual0.20.40.4
3Hmm -DimensionSequence of Observations? 1 : Dimension Verbal/Visualpicture ? text ? text? 2 : Dimension Activist/Reflectortheory ? example ?lowTypeInteractive? 1 : Dimensiontheory ? example ?LO 1picturetheorynot assignedPragmatist/TheoristlowLO 2textexamplenot assignedLO 3textnot assignedlow
4Hmm -Dimension Sequence of ObservationsSequence of State TransitionsStudent Style? 1picture ? text ? textvisual ? verbalverbal? 2theory ? example ? lowreflector ? reflector ? reflectorreflector? 1theory ? example ? lowtheorist ? theorist ? theoristtheorist
5VI.
			© 2013 Global Journals Inc. (US) Year 2013
			© 2013 Global Journals Inc. (US) Year
		
		
			

				


* 
	
		A tutorial on Hidden Markov models. Signal Processing and Artificial Neural Networks Laboratory
		
			RDugad
		
		
			UBDesai
		
		No.: SPANN-96.1
		
			1996
			Bombay
		
		
			Dept of Electrical Engineering, Indian Institute of Technology
		
	
	Technical Report




* 
	
		
		
			DunnDunn
		
		
			2003
		
	




* 
	
		The Dunn and Dunn Learning Style Model and Its Theoretical Cornerstone
		
			RitaDunn
		
		
			KennethDunn
		
		
			2003
			New York
		
		
			St John's University
		
	




* 
	
		Learning and Teaching Styles in Engineering Education
		
			RMFelder
		
		
			LKSilverman
		
	
	
		Journal of Engineering Education
		
			1988
		
	




* 
	
		The Kolb Learning Style Inventory, Version3. Boston: Hay Group
		
			DAKolb
		
		
			1999. 1992
			Mumford
		
	




* 
	
		The Manual of Learning Styles
		
			PeterHoney
		
		
			AlanMumford
		
		
			1992
			Peter Honey Publications
			Maidenhead
		
	




* 
	
		Styles and Strategies of Learning
		
			GPask
		
	
	
		British Journal of Educational Psychology
		
			1976
		
	




* 
	
		
			RRiding
		
		
			SRayner
		
		Cognitive Styles and Learning Strategies: Understanding Style Differences in Learning Behaviour
				London
		
			David Fulton Publishers Ltd
			1998
		
	




* 
	
		Explicit Intelligence in Adaptive Hypermedia: Generic Adaptation Languages for Learning Preferences and Styles
		
			NataliaStash
		
		
			AlexandraCristea
		
		
			PaulDeBra
		
	
	
		Proceedings of HT2005 CIAH Workshop
				HT2005 CIAH WorkshopSalzburg, Austria
		
			2005
		
	




* 
	
		Meta-cognitive, Cognitive and Affective Aspects of Learning Styles and Strategies: a Phenomenon graphic Analysis. Higher Education
		
			JDVermunt
		
		
			1996
		
	




* 
	
		
			HAWitkin
		
		
			CAMoore
		
		
			DRGoodenough
		
		
			PWCox
		
	
	
		Field-dependent and Field-independent Cognitive Styles and Their Educational Implications
				
			1977
		
	




* 
	
		
			ChristianWolf
		
		iWeaver: Towards Learning Style-based e-Learning in Computer Science Education. Australasian Computing Education Conference (ACE2003)
				Adelaide, Australia
		
			20
		
	
	Conferences in Research and Practice in Information Technology