Adaptive Learning Resources Sequencing in Educational Hypermedia Systems

Karampiperis, P., & Sampson, D. (2005). Adaptive Learnin Resources Sequencin in Educational Hypermedia Systems. Educational Technoloy & Society, 8 (4), 28-47. Adaptive Learnin Resources Sequencin in Educational Hypermedia Systems Pythaoras Karampiperis and Demetrios Sampson Advanced e-services for the Knowlede Society Research Unit Informatics and Telematics Institute, Centre for Research and Technoloy Hellas 42, Arkadias Street, Athens, GR-5234 Greece Department of Technoloy Education and Diital Systems University of Piraeus, 50, Androutsou Street, Piraeus GR-8534, Greece pythk@iti.r sampson@iti.r ABSTRACT Adaptive learnin resources selection and sequencin is reconized as amon the most interestin research questions in adaptive educational hypermedia systems (AEHS). In order to adaptively select and sequence learnin resources in AEHS, the definition of adaptation rules contained in the Adaptation Model, is required. Althouh, some efforts have been reported in literature aimin to support the Adaptation Model desin by providin AEHS desiners direct uidance or semi-automatic mechanisms for makin the desin process less demandin, still it requires sinificant effort to overcome the problems of inconsistency, confluence and insufficiency, introduced by the use of rules. Due to the problems of inconsistency and insufficiency of the defined rule sets in the Adaptation Model, conceptual holes can be enerated in the produced learnin resource sequences (or learnin paths). In this paper, we address the desin problem of the Adaptation Model in AEHS proposin an alternative sequencin method that, instead of eneratin the learnin path by populatin a concept sequence with available learnin resources based on pre-defined adaptation rules, it first enerates all possible learnin paths that match the learnin oal in hand, and then, adaptively selects the desired one, based on the use of a decision model that estimates the suitability of learnin resources for a tareted learner. In our simulations we compare the learnin paths enerated by the proposed methodoloy with ideal ones produced by a simulated perfect rule-based AEHS. The simulation results provide evidence that the proposed methodoloy can enerate almost accurate learnin paths avoidin the need for definin complex rule sets in the Adaptation Model of AEHS. Keywords Adaptive Educational Hypermedia, LO Sequencin, Personalization, Learnin objects. Introduction and Problem Definition elearnin can be viewed as an innovative approach for deliverin well desined, learner-centered, interactive, and facilitated learnin environment to anyone, anyplace, anytime by utilizin the attributes and resources of various diital technoloies alon with other forms of learnin materials suited for open, flexible, and distributed learnin environment, (Khan, 200). However, elearnin courses have witnessed hih drop out rates as learners become increasinly dissatisfied with courses that do not enae them (Meister, 2002; Frankola, 200). Such hih drop out rates and lack of learner satisfaction are due to the one size fits all approach that most current elearnin course developments follow (Stewart et. al., 2005), deliverin the same static learnin experience to all learners, irrespective of their prior knowlede, experience, preferences and/or learnin oals. Adaptive Educational Hypermedia (AEH) (Brusilovsky, 200; De Bra et. al., 2004) solutions have been used as possible approaches to address this dissatisfaction by attemptin to personalize the learnin experience for the learner. This learner empowerment can help to improve learner satisfaction with the learnin experience. Towards a eneral definition of an adaptive educational hypermedia system (AEHS) reflectin the current stateof-the-art, Henze and Nejdl (Henze and Nejdl, 2004) introduced a quadruple (KS, UM, OBS, AM) with the followin notation: the Knowlede Space (KS), that contains two sub-spaces. The first one, referred to as, the Media Space contains educational resources and associated descriptive information (e.. metadata attributes, usae attributes etc.) and the second, referred to as, the Domain Model contains raphs that describe the structure of the domain knowlede in-hand and the associated learnin oals. the User Model (UM), that describes information and data about an individual learner, such as knowlede status, learnin style preferences, etc. The User Model contains two distinct sub-models, one for representin the learner s state of knowlede, and another one for representin learner s conitive ISSN 436-4522 (online) and 76-3647 (print). International Forum of Educational Technoloy & Society (IFETS). The authors and the forum jointly retain the copyriht of the articles. Permission to make diital or hard copies of part or all of this work for personal or classroom use is ranted without fee provided that copies are not made or distributed for profit or commercial advantae and that copies bear the full citation on the first pae. Copyrihts for components of this work owned by others than IFETS must be honoured. Abstractin with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from the editors at kinshuk@ieee.or. 28

characteristics and learnin preferences (such as learnin style, workin memory capacity etc.). This distinction is made due to the fact that the first model (Learner Knowlede Space) can be frequently updated based on the interactions of the learner with the AEHS. On the other hand, learner s conitive characteristics and learnin preferences are more static, havin the same property values durin a sinificant time period. the Observations (OBS) which are the result of monitorin learner s interactions with the AEHS at runtime. Typical examples of such observations are: whether a user has visited a resource, the amount of time spent interactin with a iven resource, etc. Observations related with learner s behavior are used for updatin the User Model. the Adaptation Model (AM), that contains the rules for describin the runtime behavior of the AEHS. These rules contain Concept Selection Rules which are used for selectin appropriate concepts from the Domain Model to be covered, as well as, Content Selection Rules which are used for selectin appropriate resources from the Media Space. These rule sets represent the implied didactic approach of an AEHS. From the above definition, it is clear that in order to define the runtime behavior of the AEHS, the definition of how learner s characteristics influence the selection of concepts to be presented from the domain model (Concept Selection Rules), as well as the selection of appropriate resources (Content Selection Rules), is required. In the literature, there exist several approaches aimin to support the desin of these rules by providin either direct uidance to AEHS desiners, such as the Authorin Task Ontoloy (ATO) (Aroyo and Mizouchi, 2004) and the Adaptive Hypermedia Architecture (AHA) (De Bra and Calvi, 998; De Bra et. al., 2002), or semiautomatic mechanisms for makin the rule desin process less demandin, such as the Layered AHS Authorin- Model and Operators (LAOS) (Cristea and Mooij, 2003) and the Adaptive Course Construction Toolkit (ACCT) (Daer et. al., 2005). However, still the desin of adaptive educational hypermedia systems requires sinificant effort (De Bra, Aroyo and Cristea, 2004), since dependencies between educational characteristics of learnin resources and learners characteristics are too complex to exhaust all possible combinations. This complexity introduces several problems on the definition of the rules required (Wu and De Bra, 200), namely: Inconsistency, when two or more rules are conflictin. Confluence, when two or more rules are equivalent. Insufficiency, when one or more rules required have not been defined. The problems of inconsistency and insufficiency of the defined rule sets are responsible for eneratin conceptual holes to the produced learnin resource sequence (learnin path). This is due to the fact that, even if appropriate resources exist in the Media Space, the conflict between two or more rules (inconsistency problem) or the absence of a required rule (insufficiency problem), prevents the AEHS to select them and use them in the learnin resource sequence. As a result, either less appropriate resources are used from the Media Space, or required concepts are not covered at all by the resultin path. In this paper, we address the desin problem of the Adaptation Model in adaptive educational hypermedia systems proposin an alternative to the rule-based desin approach. The proposed alternative sequencin method, instead of eneratin the learnin path by populatin a concept sequence with available learnin resources based on pre-defined adaptation rules, it first enerates all possible learnin paths that match the learnin oal in hand, and then, adaptively selects the desired one, based on the use of a decision model that estimates the suitability of learnin resources for a tareted learner. This decision model mimics an instructional desiner s decision model on the selection of learnin resources (Karampiperis and Sampson, 2004). In order to evaluate the proposed sequencin methodoloy, we compare the produced learnin paths with those produced by a simulated perfect rule-based AEHS, usin a specific Domain Model and Media Space. The simulation results provide evidence that the proposed methodoloy can enerate almost accurate sequences avoidin the need for definin complex rule sets in the Adaptation Model of AEHS. The paper is structured as follows: First, we discuss the eneralized architecture of AEHS and present the abstract layers for adaptive educational hypermedia sequencin as they have been proposed in the literature. Then we present the current trends in desin tools for adaptive educational hypermedia focusin on the methods used for the definition of the Adaptation Model. In Section 3, we present our proposed methodoloy for adaptive educational hypermedia sequencin. Finally, we present simulation results of the proposed approach and discuss our findins and the conclusions that can be offered. 29

2. Adaptive Educational Hypermedia Systems: A Literature Review and Discussion Current state-of-the-art adaptive educational hypermedia systems such as AHA! (De Bra et. al., 2002), OntoAIMS (Aroyo et. al., 2003), The Personal Reader (Dolo et. al., 2004), WINDS (Kravcik and Specht, 2004), ACCT (Daer et. al., 2005) are based on the Adaptive Hypermedia Application Model (AHAM) (De Bra, Houben and Wu, 999). Fiure : Generalized Architecture of Adaptive Educational Hypermedia Systems The AHAM builds upon the Dexter model (Halasz and Schwartz, 994), that is, a common model for hypertextbased systems that was desined for eneral purpose adaptive web applications. The AHAM model refines the Dexter model so as to be used for educational purposes and extends the hypertext resources to include the full variety of hypermedia objects. The AHAM model consists of two main layers, namely, the run-time layer which contains the adaptation enine that performs the actual adaptation and the storae layer, which stores information about the Media Space, the Domain Model, the User Model and the Adaptation Model. Fiure presents a eneralized architecture of an AEHS, presentin the main components of the AHAM model and their structural interconnection. The dashed lines in this fiure represent a loical connection between the linked models. Accordin to the above architecture the desin process of an AEHS involves four key steps (Brusilovsky, 2003): Desinin the Domain Model, that is, the process of desinin a hierarchy of learnin oals, as well as, a concept hierarchy (Domain Concept Ontoloy) for describin the subject domain concepts. For each learnin oal specified in the Learnin Goals Hierarchy, a set of associated concepts in the Domain Concept Ontoloy need to be specified. This information is used by the AEHS to determine which concepts need to be covered for reachin a specific learnin oal. Desinin the User Model, that is, the process of desinin the Learner Knowlede Space, as well as, desinin the model for learner s conitive characteristics and preferences. For the desin of the Learner Knowlede Space, there exist two main approaches, the overlay modelin (Paiva and Self, 995) where the learner s state of knowlede is described as a subset of the Domain Concept Ontoloy and the stereotype modelin (Beaumont, 994) where learners are classified into stereotypes inheritin the same characteristics to all members of a certain class. Desinin the Media Space, that is, the process of desinin the educational resource description model. This model describes the educational characteristics of the learnin resources e.. the learnin resource type, or its difficulty, as well as structural relationships between learnin resources e.. if a resource requires another resource. For each learnin resource contained in the Media Space a set of related concepts from the 30

Domain Concept Ontoloy need to be specified. This information is used by the AEHS to determine if a specific learnin resource covers a certain concept of the subject domain. Desinin the Adaptation Model that is the process of definin the concept selection rules that are used for selectin from the Domain Model appropriate concepts to be covered, as well as, the content selection rules that are used for selectin appropriate resources from the Media Space. The concept selection rules are defined over the Learner Knowlede Space which represents the learner s state of knowlede by comparin it with the Domain Concept Ontoloy. The content selection rules are defined over the learner s conitive characteristics and preferences, relatin the educational characteristics of learnin resources defined in the educational resource description model with the learner s attributes in the User Model. After desinin the AEHS by followin the above mentioned steps, the adaptation enine (Adaptation Rule Parser in Fiure ), is responsible for interpretin the adaptation rules specified in the Adaptation Model in order to enerate personalized learnin paths. This process is called in the literature adaptive educational hypermedia sequencin. Followin the previous discussion on the systematic desin of AEHS, one could identify three distinct desin roles, namely: The Domain Expert, that is, the person who is responsible for definin the structure of the subject domain (Domain Concept Ontoloy), the structure of the Learner Knowlede Space, as well as, the concept selection rules of the Adaptation Model. The Instructional Desiner, that is, the person who is responsible for definin the learner conitive characteristics and preferences of the User Model, the structure of the educational resource description model, as well as, the content selection rules of the Adaptation Model. The Content Expert, that is, the person who develops the learnin resources and structures the Media Space by describin the produced learnin resources usin the educational resource description model. In practice, these distinct roles do not operate independently, but, they cooperate for desinin some of the system s models. As presented in Table, the Domain Expert and the Instructional Desiner need to work toether for the definition of the Learnin Goals Hierarchy, since learnin oals are stronly related to the concept and content selection rules. Additionally, the Instructional Desiner and the Content Expert need to work toether for the definition of the educational resource description model, since, on one hand, this model is used for describin each learnin resource developed by the Content Expert and, on the other hand, it is stronly related to the content selection rules defined by the Instructional Desiner. Desin Roles Domain Expert Instructional Desiner Content Expert Table. Role Participation in the desin of AEHS models AEHS Models Domain Model User Model Educational Adaptation Model Learnin Domain Learner Learner Resource Concept Content Goals Concept Characteristics Knowlede Description Selection Selection Hierarch Ontoloy & Preferences Space Model Rules Rules X X X X X X X X X Next section presents the current state-of-the-art tools for desinin AEHS that implement the above mentioned abstract sequencin model, focusin on the methods used for the definition of the content selection rules in the Adaptation Model. 2. Desinin methods of the Adaptation Model in AEHS Adaptive educational hypermedia sequencin is based on two main processes, namely, the concept selection process and the content selection process. In the concept selection process, a set of learnin oals from the Learnin Goals Hierarchy is selected by the learner e.. the AIMS (Aroyo and Mizouchi, 2004), or in some cases by the desiner of the AEHS e.. INSPIRE (Papanikolaou et. al., 2003). For each learnin oal, related concepts from the Domain Concept Ontoloy are selected. These concepts are filtered by the pre-existin knowlede of the learner (Learner Knowlede Space) creatin a sequence of missin concepts that need to be covered in order to reach the selected learnin oals e.. the APeLS (Conlan et. al., 2002). 3

In the content selection process, learnin resources for each concept of the concept sequence are selected from the Media Space based on the content selection rules that relate the educational characteristics of learnin resources with the conitive characteristics and learnin preferences of learners. The result of this process is a personalized learnin path that matches the selected learnin oals. Typical AEHS examples that utilize this process are the ApeLS (Conlan et. al., 2002) and the MOT (Cristea, 2004b; Cristea and Stewart, in press). Fiure 2 presents the abstract layers of adaptive educational hypermedia sequencin, demonstratin the connection of the above mentioned processes. Fiure 2: Abstraction Layers of Adaptive Educational Hypermedia Sequencin In literature, two main approaches appear to be used for the definition of the content selection rules by the AEHS Desiners Team, namely, the direct definition and the indirect definition usin predefined adaptation patterns. In the direct definition approach, the content selection rules are defined by the Instructional Desiner durin the desin process and they are based on the elements of the Resource Description Model, which is specified throuh the collaboration with the Content Expert. On the other hand, in the indirect definition approach, predefined adaptation patterns (or templates), which contain both the structure of the educational resource description model and the content selection rules of the Adaptation Model, are selected by the Instructional Desiner. Consequently, we can classify the desin tools for AEHS recorded in the literature, with reard to their approach for definin the content selection rules, in the followin two classes: Desin Tools supportin the direct definition of the Content Selection Rules. These systems support the Instructional Desiner in the process of directly definin content selection rules. They require the Instructional Desiner to have ood knowlede of the parameters of the system that can be adapted, as well as the details of the User Model. Typical examples of these systems are the AHA! (De Bra and Calvi, 998; De Bra et. al., 2002), the OntoAIMS (Aroyo et. al., 2003), the AIMS (Aroyo and Mizouchi, 2004), The Personal Reader (Dolo et. al., 2004), and others. Althouh these systems provide raphical environments for the definition of the content selection rules and/or visual representation of the resultin learnin/teachin scenario, still it is difficult for Instructional Desiners to overcome the problems of inconsistency, confluence and/or insufficiency of the selection rules (De Bra, Aroyo and Cristea, 2004). This is due to the fact that, on one hand, dependencies between educational characteristics of learnin resources and conitive characteristics of learners are rather complex (Cherniavsky and Soloway, 2002; Karampiperis and Sampson, 2004), and on the other hand, it is difficult for an Instructional Desiner to know the details of each User Model in use and the correspondin meaninful pedaoical adaptations required (Cherniavsky and Soloway, 2002), since there exist several different models for each learner conitive characteristic. For example, only in the case that learnin styles are used as the main adaptation parameter, there exist more than seventy different models in use (Brown et. al., 2005). 32

Desin Tools supportin the indirect definition of the Content Selection Rules. These systems use preexistin adaptation patterns (or templates) that have been a-priori defined by an Instructional Desiner durin the development phase of the desin tool. Typical examples of these systems are the MOT (Cristea, 2004b; Cristea and Stewart, in press), the ACCT (Daer et. al., 2005), and others. The main advantae of these systems is that they simplify the desin process of adaptive hypermedia, since the educational resource description model and partly the Adaptation Model is predefined. However, when more than one patterns are to be used the AEHS Desiners Team is required to know the details of each selected pattern in order to avoid the problems of inconsistency and/or confluence. Additionally, it is nearly impossible for the AEHS Desiners Team to extend an existin adaptation pattern, since the definition of new adaptation rules in a pattern would require the AEHS Desiners Team to be familiar with the implementation details of the pattern notation lanuae used. 3. The proposed Adaptive Sequencin Methodoloy As described in section 2, existin adaptive educational hypermedia systems implement a rule-based sequencin approach based on a two steps procedure. They first enerate a sequence of concepts that matches the learnin oal in hand, and then select learnin recourses for each concept of the concept sequence. Due to the problems of inconsistency and insufficiency of the defined rule sets in the Adaptation Model, conceptual holes can be enerated in the produced learnin resource sequence. To overcome this problem, we propose an alternative sequencin method that instead of eneratin the learnin path by populatin the concept sequence with available learnin resources, it first enerates all possible sequences that match the learnin oal in hand and then adaptively selects the desired personalized learnin path from the set of available paths. More precisely, the followin two steps procedure is used: Step: Learnin Paths Generation At this step a raph containin all possible learnin paths based on the relation between the Learnin Goals Hierarchy, the concepts of the Domain Concept Ontoloy and the learnin resources contained in the Media Space, is enerated. This raph is constructed as follows: Fiure 3: The proposed Abstraction Layers of Adaptive Educational Hypermedia Sequencin 33

Stepa: Construction of the Concepts Path Graph. The Concepts Path Graph (CPF) is a directed raph which represents the structure of the concepts of the Domain Concept Ontoloy that matches the learnin oal in hand. The concepts contained in the CPF are selected based on the connection between the Learnin Goals Hierarchy and the Domain Concept Ontoloy. The structure of the CPF is directly inherited by the structure of the Domain Concept Ontoloy. CPF is a simple directed raph, that is, a directed raph havin no multiple nodes. This means that each concept is contained only once in the CPF. Additionally, CPF is an acyclic directed raph, that is, a directed raph containin no directed cycles. This means that in every possible concept sequence represented by the CPF, each concept has a unique existence. Stepb: Construction of the Learnin Paths Graph. The Learnin Paths Graph (LPG) is a directed raph which represents all possible learnin paths (sequence of learnin resources) that matches the learnin oal in hand. To construct the LPG, for each concept of the CPF related learnin resources are selected from the Media Space based on the connection between the Domain Concept Ontoloy and the Resource Description Model. Each node in the CPF is then replaced by the related set of learnin resources retrieved from the Media Space. The structure of the learnin resources set is directly inherited by the structure of the Media Space. The final raph is the Learnin Paths Graph. Assumin that the Media Space does not contain circular references between learnin resources, the LPG is aain a simple acyclic directed raph. Althouh this assumption does not directly affect either the desin of an AEHS, nor our sequencin methodoloy, it is necessary for avoidin infinite learnin paths. Step2: Personalized Learnin Path Selection. At this step a personalized learnin path is selected from the raph that contains all the available learnin paths based on learner s attributes in the User Model. As a result, we introduce an additional layer (Fiure 3) in the abstract sequencin layers of adaptive educational hypermedia systems, namely the Learner Adaptation Layer, which is used for selectin the personalized learnin path. In the proposed sequencin method, we replace the content selection rules defined in the Adaptation Model with a decision-makin function that estimates the suitability of a learnin resource for a specific learner by relatin the educational characteristics of learnin resources defined in the educational resource description model with the learner s conitive characteristics and preferences stored in the User Model. This suitability function is used for weihtin each connection of the Learnin Paths Graph. From the weihted raph, we then select the most appropriate learnin path for a specific learner (personalized learnin path) by usin a shortest path alorithm. Next sections present the methodoloy used for creatin the suitability function, as well as, for selectin the personalized learnin path for a learner. 3.. Creatin the Suitability Function Next, we present the alorithm for creatin a suitability function that estimates the suitability of a learnin object for a specific learner. In our previous work, we have proposed a decision model that constructs a suitability function which maps learnin object characteristics over learner characteristics and vice versa. We have already used that model for the direct selection of learnin objects, provin that this suitability function can safely extract dependencies between a User Model and a Resource Description Model (Karampiperis and Sampson, 2004). In that work the User Model elements were not directly defined by the Instructional Desiner, but they were dynamically selected from a set of elements durin the suitability calculation phase. In this paper, we construct a similar suitability function with the assumption that the elements of the User Model are directly defined by the Instructional Desiner and remain the same durin the whole life cycle of the AEHS. To this end, before proceedin with the calculation of the suitability function, we assume that the learners conitive characteristics and preferences stored in the User Model, as well as, the structure of the Educational Resource Description Model have already been defined by the Instructional Desiner. The process of creatin the suitability function consists of the followin steps, as shown in Fiure 4: Step: Reference Sets Generation The first step of the suitability calculation process includes the eneration of the reference sets of learnin objects and learners that will be used for calculatin the suitability function. More precisely, we enerate two sets of learnin objects, namely, the Learnin Objects Trainin Set (LOTS) and the Learnin Objects Generalization Set (LOGS), as well as, two sets of learners, namely, the Learners Trainin Set (LTS) and the Learners Generalization Set (LGS). The two trainin sets (LOTS and LTS) are used for calculatin the suitability function, and the two eneralization sets (LOGS and LGS) are used for evaluatin the consistency of the produced suitability function. 34

Each one of the enerated reference learnin objects has a unique identifier of the form LO i and is characterized by a set of n independent properties = (, 2,, n ) of the Educational Resource Description Model. Similarly, each one of the enerated reference learners has a unique identifier of the form L j and is characterized by a set of m independent properties L j L j L j L j u = ( u, u2,, um ) of the User Model. The reference learnin objects are randomly enerated with normal distribution over the value space of each metadata element of the Resource Description Model. Similarly, the reference learners are randomly enerated with normal distribution over the value space of each learner characteristic of the User Model. Step2: Reference LO ratin by the Instructional Desiner For each reference learner L j contained in the LTS, we ask the Instructional Desiner to define his/her preference ratin of the reference learnin objects contained in LOTS, as well as, to define his/her preference ratin of the reference learnin objects contained in LOGS. These preference ratins are expressed usin two preference relations, namely, the strict preference relation and the indifference relation. A strict preference relation means that a learnin object is preferred from another one and an indifference relation means that two learnin objects are equally preferred. Additionally, for each reference learner L j contained in the LGS, we ask the Instructional Desiner to define his/her preference ratin of the reference learnin objects contained in LOGS. START Reference Set Generation Reference Set of Learnin Objects (Trainin & Generalization Set) Reference Set of Learners (Trainin & Generalization Set) Step Expression of Instructional Desiner s Reference LO ratin on the Reference Set of Learners Add an LO Instance to the Trainin Set Step 2 Suitability Function Parameters Calculation Step 3 Fail Consistency Check based on Learner Trainin Set Pass Extrapolation on the entire set of Learner Instances Add a Learner Instance to the Trainin Set Consistency Check based on Learner Generalization Set Pass Fail Step 4 END Fiure 4: Suitability Function Creation Workflow 35

Step3: Suitability Function Parameters Calculation For a specific learner L j we define as marinal suitability function of the Resource Description Model property k a function that indicates how important is a specific value of the property k when calculatin the suitability of a learnin resource LO i for the learner L j. This function has the followin form (Karampiperis and Sampson, 2004): L 2 j LO L j L i j LO L i j s ( ) exp( ) = a + b k c, where k is the property value of learnin object L j L j L j LO i in the k element of the Resource Description Model and a R, b R c R k, k are k parameters that define the form of the marinal suitability function. The calculation of these parameters for all k properties of the Resource Description Model lead to the calculation of the suitability function for the learner L j. More precisely, for a specific learner L j we define the suitability function as the areation of the marinal suitability functions for the learner Lj, as follows: n L LO j L i j S ( ) = s ( ) with the followin additional notation: n k = L j s ( ) : Marinal suitability of the k element of the Resource Description Model, valued k for the learnin object LO i, L j LO S i : The lobal suitability of the learnin object LO i for the learner L j. ( ) Lj Lj If S LO is the lobal suitability of a learnin object LO and SLO 2 is the lobal suitability of a learnin object LO 2 for the learner L j, then the followin properties enerally hold for the suitability function S: Lj Lj S LO > S ( ) ( 2 ) LO LO P LO 2, Lj Lj S LO = S ( LO ) I( LO2 ) LO 2 where P is the strict preference relation and I the indifference relation in Instructional Desiner s preference ratin. These properties express that for a specific learner L j, when a learnin object LO is preferred from another learnin object LO 2, then the suitability function for LO is reater than the suitability function for LO 2 and vise versa. Similarly, when two learnin objects LO and LO 2 have the same preference ratin for a specific learner L j, then they also have the same suitability function value. Usin the provided by the Instructional Desiner preference ratin of the reference learnin objects contained in LOTS, for each reference learner L j contained in the LTS, we define the suitability L j L j L j L j differences Δ = ( Δ, Δ2,, Δq ) for the reference learner L j, where q is the number of learnin objects in the LOTS and Δ L j = L j L j l S LO S 0 l LO the suitability difference between two subsequent l+ learnin objects in the rated LOTS. We then define an error function e for each suitability difference: L j L j L j L j Δ l = S LO S + 0 l LO e l+ l. Usin Larane multipliers and Conjuate Gradient, we can then solve for each one of the learner instances L j in the LTS the followin constrained optimization problem: q Minimize L Δ j 2 l > 0 if (LOl ) P( LOl + ) ( el ) subject to the constraints: and l= Δ l = 0 if (LOl ) I( LOl+ ) L j 0 s ( k ), k k This optimization problem leads to the calculation of the values of the parameters a, b and c for each k property of the Resource Description Model over the instances of the LTS, that is, for each separate learner profile included in the LTS. Step4: Consistency Check and Extrapolation We then evaluate the consistency of the resultin suitability function, that is, the evaluation of how well the suitability function works for learnin objects and/or learners that have not been used in the suitability function parameters calculation (step 3). To this end, we first use the provided by the Instructional Desiner preference ratin of the reference learnin objects contained in LOGS, for each reference learner L j contained in the LTS. 36

For a reference learner L j, we estimate usin the suitability function calculated in the previous step (step 3) the Instructional Desiner s preference ratin of each learnin object contained in LOGS. We then compare the provided by the Instructional Desiner preference ratin with the estimated one. If the preference ratin estimation of a learnin object LO i in LOGS is different than that provided by the Instructional Desiner, we add the learnin LO i in the Learnin Object Trainin Set (LOTS) and recalculate the suitability function parameters (step 3). If the estimated and the provided preference ratins are the same, then we eneralize the resulted suitability function from the LTS to all learners, by calculatin the correspondin suitability values for L every learner property u j z, usin the followin linear interpolation formula: L L L2 s ( ) ( ) = ( ), if s s L j s ( ) = L j L L LO u i z u, where z L2 L L2 L2 s ( ) + [ ( ) ( )] ( ) > ( ) s L L s, if s s 2 u z u z L L and L 2 are the learners of the LTS closest (measured by Euclidean distance) to the learner L j, u z L2 L L2 and u z are the values of learner property u z for learners L and L 2 respectively, and s k and s k are the marinal suitability functions of the Resource Description Model property k for learners L and L 2 respectively. After the extrapolation on the entire set of learner instances, we evaluate aain the consistency of the resultin suitability function, usin the provided by the Instructional Desiner preference ratin of the reference learnin objects contained in LOGS, for each reference learner L j contained in the LGS. For a reference learner L j, we estimate usin the suitability function calculated in the previous step (step 3) the Instructional Desiner s preference ratin of each learnin object contained in LOGS. We then compare the provided by the Instructional Desiner preference ratin with the estimated one. If the preference ratin estimation for a learner L j in LGS is different than that provided by the Instructional Desiner, we add the learner L j in the Learners Trainin Set (LTS) and recalculate the suitability function parameters (step 3). 3.2. Selectin a Personalized Learnin Path Followin our proposed 2-step methodoloy for adaptive sequencin, in order to be able to select from the Learnin Paths Graph the learnin path that matches the characteristics and preferences of a specific learner, we need to add learner-related information to the LPG. This information has the form of weihts on each connection of the LPG and represents the inverse of the suitability of a learnin resource for the specific learner. This means that the hiher value a weiht in the LPG has, the less suitable the correspondin learnin object in the sequence is for a specific learner. For a specific learner Lj we define the weihtin function for each directed connection L j LO L i j LO L i (ede) of the Learnin Paths Graph as W ( ) = S ( ) [0,], where j LO S ( i ) is the lobal suitability for the learner L j of the tareted learnin object LO i. in the ede. After weihtin the LPG usin the weihtin function, we need to find the most appropriate learnin path for a learner. Since the weihts in the LPG are calculated in such a way that the lower value they have the more suitable a learnin object is, the calculation of the most appropriate learnin path is equivalent to the calculation of the shortest path in the LPG. By relaxin the edes of the LPG accordin to a topoloical sort of its vertices (nodes of the raph), we can compute the shortest path. The alorithm starts by topoloically sortin the LPG to impose a linear orderin on the vertices. If there is a path from vertex u to vertex υ, then u precedes υ in the topoloical sort (Fiure 5a). Let us call V the set of vertices contained in the LPG. For each vertex υ V, we maintain an attribute d[υ] called shortest-path estimation, which is an upper bound on the weiht of a shortest path from source s to υ. Additionally, for each vertex υ V, we maintain an attribute π[υ] called shortest-path predecessor. We initialize the shortest-path estimates and predecessors usin the followin values: π[υ]=nil for all υ V, d[s]=0, and d[υ]= for υ V {s} (Fiure 5a). We make just one pass over the vertices in the topoloically sorted order. As we process each vertex, we relax each ede that leaves the vertex. The process of relaxin an ede (u,υ) consists 37

of testin whether we can improve the shortest path to υ found so far by oin throuh u and, if so, updatin d[υ] and π[υ]. A relaxation step may decrease the value of the shortest-path estimate d[υ] and update υ s predecessor field π[υ] (Fiure 5b-). Fiure 5: The execution of the alorithm for personalized learnin path selection from the LPG. The d values are shown within the vertices, and shaded edes indicate the π values. The result of this process is the calculation of the shortest path in the LPG that corresponds to the sequence of learnin objects that are most suitable for a specific learner L j. 4. Settin up the Simulation In this paper we present a sequencin methodoloy for AEHS that aims to overcome the problem of eneratin sequences with conceptual holes. In order to evaluate the proposed sequencin methodoloy, we compare the produced learnin paths with those produced by a simulated perfect rule-based AEHS, usin a specific Domain Model and Media Space. A perfect rule-based AEHS is assumed to contain consistent and sufficient adaptation rule sets. As a result, it is anticipated that such a system would enerate for a desired learnin oal, solid learnin paths with no conceptual holes. For the simulation of the learnin paths produced by a perfect rule-based AEHS, we use a specific Domain Model and Media Space (as described later in this section) and enerate for each learnin oal specified in the Learnin Goals Hierarchy all consistent learnin paths that can be defined over the specific Media Space. In our simulations we measure how close the learnin paths produced by our proposed methodoloy are to these ideal paths. By this way, we intent to demonstrate the capacity of the proposed methodoloy and investiate parameters that influence this performance. In order to setup our simulations, we use the common desin steps of an AEHS, as described in section 2. More specifically: Desinin the Domain Model. The selected domain for our simulations was the Computer Science Domain. For the description of the subject domain concepts, that is, the Domain Concept Ontoloy, we extracted the ontoloy from the ACM Computin Curricula 200 for Computer Science (ACM, 200). As discussed in section 2, the use of ontoloies for structurin the Domain Concept Ontoloy is commonly used in AEHS, since it provides a 38

standard-based way for knowlede representation (Henze, Dolo and Nejdl, 2004; Aroyo and Dicheva, 2004). The extracted ontoloy is complete consistin of 950 topics oranized in 32 units and 4 areas (see Table 2). A partial view of the concept hierarchy in the domain ontoloy in use is shown in Fiure 6. Computer Science. Consists of 2. Similar to 3. Opposite of 4. Related with Concept Relation Classes Intellient Systems Software Enineerin Knowlede representation and reasonin Machine learnin and neural networks Natural lanuae processin Software Desin Software Validation Supervised learnin 3 3 Unsupervised learnin 2 Self-Oranized learnin Validation Plannin Testin Fundamentals Object-Oriented Testin Back-Propaation Support Vector Machines Reinforcement learnin 4 Dynamic Prorammin Fiure 6: Partial View of Concept Hierarchy in the Domain Concept Ontoloy in use (ACM Computin Curricula 200 for Computer Science) For the description of the relations between the subject domain concepts we used four classes of concept relationships, as shown in Fiure 6, namely: Consists of, this class relates a concept with its sub-concepts Similar to, this class relates two concepts with the same semantic meanin Opposite of, this class relates a concept with another concept semantically opposite from the oriinal one Related with, this class relates concepts that have a relation different from the above mentioned Table 2: Subject Domain Concepts covered in the Ontoloy Area Units Topics Discrete Structures 6 45 Prorammin Fundamentals 5 32 Alorithms and Complexity 7 Architecture and Oranization 9 55 Operatin Systems 2 7 Net-Centric Computin 9 79 prorammin lanuaes 75 Human-Computer Interaction 8 47 Graphics and Visual Computin 84 Intellient Systems 0 06 Information Manaement 4 93 Social and Professional Issues 0 46 Software Enineerin 2 85 Computational Science 4 6 Furthermore, for the definition of the Learnin Goals Hierarchy in our simulations, we have used aain the ACM Computin Curricula 200 for Computer Science, which defines for each subject domain concept associated learnin objectives (ACM, 200). From this list of learnin objectives we have created a Learnin Goals Hierarchy which is presented in Fiure 7. We then associated each topic of the 950 topics included in the Domain Concept Ontoloy in use with at least one node of the enerated Learnin Goals Hierarchy, so as to provide a connection between learnin oals and concepts of the particular Domain Concept Ontoloy in hand. 39

Desinin the User Model. For the desin of the User Model in our simulations, we have used an overlay model for representin the Learners Knowlede Space and a stereotype model for representin learners preferences. More precisely, for the learners knowlede level we track the existence of a related certification for each node of the Learners Knowlede Space, the evaluation score in testin records and the number of attempts made on the evaluation. For modelin of learners preferences we use learnin styles accordin to Honey and Mumford model (Honey and Mumford, 992), as well as modality preference information consistin of three modality types, namely, the visual modality, the textual modality, the auditory modality and the mixed modality preferences. Each element of the User Model was mapped to the IMS Learner Information Packae (IMS LIP) specification, as shown in Table 3. Fiure 7: Learnin Goals Hierarchy (ACM Computin Curricula 200 for Computer Science) Table 3: Usin the IMS LIP specification for representin User Model elements User Model Element IMS LIP Element Explanation Learnin Style Accessibility/Preference/typename The type of conitive preference Accessibility/Preference/prefcode The codin assined to the preference Modality Preference AccessForAll/Context/Content The type of modality preference QCL/Level The level/rade of the QCL Knowlede Level Activity/Evaluation/noofattempts The number of attempts made on the evaluation. Activity/Evaluation/result/interpretscope Information that describes the scorin data Activity/Evaluation/result/score The scorin data itself. Desinin the Media Space. For the desin of the Media Space in our simulations we have used as Educational Resource Description Model a subset of the IEEE Learnin Object Metadata standard elements, illustrated in Table 4. The Areation Level and the Relation/Kind elements are used for structurin the Media Space and the Classification element is used for connectin learnin resources with the concepts of the Domain Concept Ontoloy. The Areation Level was used for classifyin the available learnin resources in two classes, namely, the raw media and the structured learnin objects (Table 5). Each learnin resource was taed with a unique identifier dependin on the areation level class that it belons. For example, the identifier of learnin resources with areation level has the form of AG:, whereas, the identifier of learnin resources with areation level 2 has the form of AG2:LOj, where i and j are the unique identifiers of the learnin resources inside a specific areation class. In order to define the structure of learnin resources at areation level 2 (that is, a collection of several learnin resources at areation level ) we have used the Relation Cateory of the IEEE LOM standard. 40

More specifically, in our simulations we have used eiht types of relationships out the 2 predefined values at the Dublin Core Element Set (DCMI, 2004), namely: is part of / has part references / is referenced by is based on / is basis for requires / is required by IEEE LOM Cateory General Educational Table 4: Educational Resource Description Model used IEEE LOM Element Structure Areation Level Interactivity Type Interactivity Level Semantic Density Typical Ae Rane Difficulty Intended End User Role Context Typical Learnin Time Learnin Resource Type Explanation Underlyin oranizational structure of a Learnin Object The functional ranularity of a Learnin Object Predominant mode of learnin supported by a Learnin Object The deree to which a learner can influence the aspect or behavior of a Learnin Object. The deree of conciseness of a Learnin Object Developmental ae of the typical intended user. How hard it is to work with or throuh a Learnin Object for the typical intended taret audience. Principal user(s) for which a Learnin Object was desined, most dominant first. The principal environment within which the learnin and use of a LO is intended to take place. Typical time it takes to work with or throuh a LO for the typical intended taret audience. Specific kind of Learnin Object. The most dominant kind shall be first. Relation Kind Nature of the relationship between two Learnin Objects Classification Taxon Path A taxonomic path in a specific classification system. Table 5: Learnin Objects Areation Level accordin to IEEE LOM standard IEEE LOM Element Value Space Description The smallest level of areation, e.. raw media data or framents General/Areation_Level A collection of level learnin objects, e.. 2 a lesson chapter or a full lesson A partial view of the Media Space based on the use of the IEEE LOM Areation Level element and the Relation/Kind element is presented in Fiure 8. Furthermore, for each learnin resource included in the Media Space, a set of related concepts from the Domain Concept Ontoloy is specified usin the Classification element of the IEEE LOM standard. This element describes the position of a specific learnin object within a particular classification system and it is typically used in AEHS to determine if a specific learnin resource covers a certain concept of the subject domain. Typical systems that used this approach are the Personal Reader (Dolo et. al., 2004), the WINDS (Kravcik and Specht, 2004) and others. In the literature, several approaches exist that interate the IEEE LOM metadata elements within domain concept ontoloies (Kay and Holden, 2002; Sicilia et. al., 2004; Hayashi, Ikeda and Mizouchi, 2004; Simon et. al., 2004). The use of the classification element of the IEEE LOM standard, on one hand, models the connection between concepts of the Domain Concept Ontoloy and the learnin resources, and on the other hand, enables the separation of the Educational Resource Description Model from the Domain Concept Ontoloy. This separation enables the use of separate metadata records for learnin resources, thus, enablin the use of resources and associated metadata contained in external from the AEHS repositories. Desinin the Adaptation Model. For the desin of the Adaptation Model in our simulations, we have used the methodoloy presented in section 3 based on a set of 50 learnin object metadata records (30 for the Learnin Objects Trainin Set and 20 for the Learnin Objects Generalization Set) and a set of 0 simulated learner 4

instances (5 for the Learners Trainin Set and 5 for the Learners Generalization Set). These sets were used to calculate the suitability function presented in section 3. of this paper. Fiure 8: Partial View of Media Space Representation For our simulations we created an additional set of learnin object metadata records that we call Learnin Objects Estimation Set - LOES, consistin of 42.500 records (that is, 50 simulated learnin objects for each one of the 950 topics), with normal distribution over the value space of each metadata element. Additionally, we created a set of 20 simulated learner instances that we call Learner Estimation Set - LES, with normal distribution over the value space of each learner characteristic. These estimation sets were used for evaluatin the efficiency of the proposed approach in eneratin learnin paths with no conceptual holes, as it is discussed in the next section. 5. Simulation Results and Discussion In our simulations we evaluate the proposed sequencin methodoloy by comparin the produced learnin paths with those produced by a simulated perfect rule-based AEHS, as described in the previous section. To this end, we have defined an evaluation criterion based on Kendall s Tau (Wilkie, 980), which measures the match between two learnin object sequences, as follows: Success (%) = 00 * + 2 N concordant k( k ) N discordant, k = n topic level where N concordant stands for the concordant pairs of learnin objects and N discordant stands for the discordant pairs when comparin the resultin learnin objects sequence with the ideal reference one and n is the maximum requested number of learnin objects per concept level. The efficiency of the proposed method was evaluated by comparin the resultin learnin object sequences with ideal reference sequences for 50 different cases (0 randomly selected learner instances from the Learner Estimation Set per level of sequence root) over the concept hierarchy. Averae evaluation results are shown in Fiure 9 presentin the success of the proposed sequencin method for different cases of sequence roots (that is, 42

the concept level in Domain Ontoloy) and different cases of maximum requested number of learnin objects per concept level (n). In this fiure the different concept levels express the depth in the Domain Ontoloy of the root concept in the desired learnin path. For example, topic levels (-5) correspond to concepts in the Domain Ontoloy with depth between one and five. These concepts are included in a Unit (see also Table 2) and they possibly include topics with depth reater than five, dependin on the structure of the Domain Ontoloy. 00 Averae Sequence Generation Success per Level of Sequence Root %Success 95 90 85 80 75 70 maxlos/concept Level n=5 n=0 n=20 n=50 65 Area Units Topic Levels (-5) Topic Levels (6-0) Concept Level in Domain Ontoloy Topic Levels (-5) Fiure 9: Averae Simulation Results for Learnin Path Selection From these results we conclude that the success rate of the resultin learnin object sequences is dependin on the concept levels that the end sequence covers, as well as the maximum requested resources for each level. The less number of resources per level are requested, the smallest would be the resulted LO sequence, producin less probability of possible mismatches. Accordinly, for the same number of requested objects per level, the hiher level the sequence root is, the loner would be the resulted sequence introducin more mismatches from the Learnin Paths Graph weihtin process. These observations introduce two main desin principles that should be followed in order to successfully enerate personalized learnin paths, namely: The Content Expert of an AEHS should desin the Media Space by creatin structured learnin resources (with Areation Level equal to 2) rather than raw media. This internal structurin, on one hand, enables the AEHS to select less (but more areated) learnin resources, and on the other hand, increases the probability of eneratin meaninful learnin paths since less decisions about the structurin of the learnin resources are taken by the AEHS. The end-user of an AEHS should request an adaptive web-based course coverin the minimum needed parts of the Domain Concept Ontoloy, in order for avoidin the eneration of hue sequences that introduce mismatches. In order to investiate in more detail these mismatches, we have desined another evaluation criterion, which measures the success in selectin appropriate resources per concept node, defined by: Selection Success (%) = 00 * Correct Learnin Objects Selected where m is the number of requested learnin objects from the Media Space per concept node. m We have evaluated the selection success on two different sub sets of the Learnin Objects Estimation Set. The first data sets contains learnin object metadata records with areation level (raw media) and the second data set contains learnin object metadata records with areation level 2 (structured learnin objects), as defined in section 4. Fiure 0 presents averae simulation results for learnin objects selection. 43

From these results we can once aain confirm the observation that usin structured learnin objects rather than raw media, increases the probability of eneratin flawless learnin paths. More analysis on the results, presented in Fiure 0, shows that when the desired number of learnin objects (m) is relatively small (less or equal to 0), the efficiency of selection is almost the same for raw media and structured learnin objects. However, when the desired number of learnin objects is relatively lare (more than 0) the success in selectin learnin objects is stronly affected by the areation level of the learnin objects. % LO Selection Success 00 98 96 94 92 90 88 86 84 82 5 0 20 50 00 A. Level 99.7 97.2 94.3 9.6 88.7 A. Level 2 00.0 98.9 97.4 95.8 94. Num of LOs/Concept Node (m) Fiure 0: Averae Simulation Results for Learnin Objects Selection If we consider that for one learner instance, the different combinations of learnin objects, calculated as the multiplication of the value instances of characteristics presented in Table 4, leads to more than one million learnin objects, it is evident that it is almost unrealistic to assume that an instructional desiner can manually define the full set of selection rules which correspond to the dependencies extracted by the proposed method and at the same time to avoid the inconsistencies, confluence and insufficiency of the produced selection rules. The simulation results demonstrate that the proposed approach is capable of extractin dependencies between learnin object and learner characteristics producin almost accurate sequences of learnin objects (that is, almost similar to the ideal ones). Furthermore, it was exhibited that the ranularity of learnin object sequences, as well as, the areation level of the learnin objects are the main parameters affectin the sequencin success. A learnin path that covers a whole concept area is more likely to produce mismatches when comparin with a sequence that covers only a specific unit or even a specific topic, and a sequence that uses raw media is more likely to produce mismatches when comparin with a sequence that uses structured learnin objects. This is due to the fact that structured learnin objects partly contain information about the underlyin pedaoical scenario. When only raw media are used for sequencin, then the pedaoical scenario is totally implied in the decisions made by the AEHS. Our future work, focuses on separatin the learnin scenario from the adaptation decision model. By this way, we anticipate, on one hand, to support better the sequencin of unstructured raw media, and on the other hand, to facilitate the support of different pedaoical strateies without redesinin the adaptation decision model. 6. Conclusions In this paper, we address the desin problem of the Adaptation Model in AEHS proposin an alternative sequencin method that instead of eneratin the learnin path by populatin a concept sequence with available learnin resources based on adaptation rules, it first enerates all possible sequences that match the learnin oal in hand and then adaptively selects the desired sequence, based on the use of a decision model that estimates the suitability of learnin resources for a tareted learner. In our simulations we compare the produced sequences by the proposed methodoloy with ideal sequences produced by a perfect rule-based AEHS. The simulation results provide evidence that the proposed methodoloy can enerate almost accurate sequences avoidin the need for 44

definin complex rule sets in the Adaptation Model of AEHS. Additionally, the simulation results showed that the success in learnin object sequencin is stronly affected by the areation level of the learnin objects and the number of the concepts covered by the desired learnin object sequence. 7. Acknowledements The work presented in this paper is partially supported by European Community under the FP6 Information Society Technoloies (IST) proramme ICLASS contract IST-507922, and the Leonardo da Vinci proramme eaccess contract EL/2003/B/F/48233. 8. References ACM. (200). ACM Computin Curricula 200 for Computer Science, Retrieved October 25, 2005, from http://www.acm.or/education/curricula.html. Aroyo, L., Mizouchi, R., & Tzolov, C. (2003). OntoAIMS: Ontoloical Approach to Courseware Authorin. Paper presented at the International Conference on Computers in Education (ICCE2003), December 2-5, 2003, Hon Kon. Aroyo, L., & Dicheva, D. (2004). The New Challenes for E-learnin: The Educational Semantic Web. Educational Technoloy & Society, 7 (4), 59-69. Aroyo, L., & Mizouchi, R. (2004). Towards Evolutional Authorin Support Systems. Journal of Interactive Learnin Research, 5 (4), 365-387. Beaumont, I. (994). User modelin in the interactive anatomy tutorin system ANATOMTUTOR. International Journal of User Modelin and User-Adapted Interaction, 4 (), 2-45. Brown, E., Cristea, A., Stewart, C., & Brailsford, T. (2005). Patterns in Authorin of Adaptive Educational Hypermedia: A Taxonomy of Learnin Styles. Education Technoloy & Society, 8 (3), 77-90. Brusilovsky, P. (200). Adaptive hypermedia. Methods and techniques of adaptive hypermedia. International Journal of User Modelin and User-Adapted Interaction, (/2), 87-0. Brusilovsky, P. (2003). Developin adaptive educational hypermedia systems: From desin models to authorin tools. In Murray, T., Blessin, S., Ainsworth, S. (Eds.), Authorin Tools for Advanced Technoloy Learnin Environment, Dordrecht: Kluwer Academic Publishers, 377-409. Cherniavsky, J. C., & Soloway, E. (2002) A survey of research questions for intellient information systems in education. Journal of Intellient Information Systems, 8 (), 5-4. Conlan, O., Hockemeyer, C., Wade, V., & Albert, D. (2002). Metadata Driven Approaches to Facilitate Adaptivity in Personalized elearnin Systems. Journal of the Japanese Society for Information and Systems in Education, (), 38-44. Cristea, A., & Mooij de A. (2003). LAOS: Layered WWW AHS Authorin Model and their correspondin Alebraic Operators Paper presented at the 2 th International World Wide Web Conference, May 20-24, 2003, Budapest, Hunary. Cristea, A. (2004a). What can the Semantic Web do for Adaptive Educational Hypermedia? Educational Technoloy & Society, 7 (4), 40-58. Cristea, A. (2004b). Is Semi-Automatic Authorin of Adaptive Educational Hypermedia possible? Advanced Technoloy for Learnin Journal, (4), 227-236. Cristea, A., & Stewart, C. (In Press). Automatic Authorin of Adaptive Educational Hypermedia. In Ma, Z. (Eds.), Web-Based Intellient e-learnin Systems: Technoloies and Applications, IDEA Publishin roup. 45

Daer, D., Wade, V., & Conlan O. (2005). Personalisation for All: Makin Adaptive Course Composition Easy. Education Technoloy & Society, 8 (3), 9-25. DCMI (2004) Dublin Core Metadata Element Set, Version., Retrieved October 25, 2005, from, http://dublincore.or/documents/dces/. De Bra, P., Houben, G. J., & Wu, H. (999). AHAM: A Dexter based Reference Model for Adaptive Hypermedia. Paper presented at the ACM Conference on Hypertext and Hypermedia, February 2-25, 999, Darmstadt, Germany. De Bra, P., & Calvi, L. (998). AHA! An open Adaptive Hypermedia Architecture. The New Review of Hypermedia and Multimedia, 4, 5-39. De Bra, P., Aerts, A., Smits, D., & Stash, N. (2002). AHA! Version 2.0, More Adaptation Flexibility for Authors. Paper presented at the World Conference on e-learnin in Corporate, Government, Healthcare & Hiher Education (ELearn'2002), October 5-9, 2002, Montreal, Canada. De Bra, P., Aroyo, L., & Cristea, A. (2004). Adaptive Web-based Educational Hypermedia. In Levene, M. & Poulovassilis, A. (Eds.), Web Dynamics, Adaptive to Chane in Content, Size, Topoloy and Use, Heidelber, Germany: Spriner, 387-40. De Bra, P., Aroyo, L., & Chepein, V. (2004). The Next Bi Thin: Adaptive Web-Based Systems, Journal of Diital Information, 5 (), Retrieved October 25, 2005, from, http://jodi.tamu.edu/articles/v05/i0/debra/. Dolo, P., Henze, N., Nejdl, W., & Sintek, M. (2004). The Personal Reader: Personalizin and Enrichin Learnin Resources Usin Semantic Web Technoloies. Paper presented at the Third International Adaptive Hypermedia and Adaptive Web-based Systems Conference (AH2004), Auust 23-26, 2004, Eindhoven, The Netherlands. Frankola, K. (200). Why online learners dropout. Workforce, 0, 53-63. Halasz, F., & Schwartz, M. (994). The Dexter Hypertext Reference Model. Communications of the ACM, 37 (2), 30-39. Hayashi, Y., Ikeda, M., & Mizouchi, R. (2004). A Desin Environment to Articulate Desin Intention of Learnin Contents. International Journal of Continuin Enineerin Education and Life Lon Learnin, 4 (3), 276-296. Henze, N., & Nejdl, W. (2004). A Loical Characterization of Adaptive Educational Hypermedia. New Review of Hypermedia and Multimedia (NRHM), 0 (), 77-3. Henze, N., Dolo, P., & Nejdl, W. (2004). Reasonin and Ontoloies for Personalized E-Learnin. Educational Technoloy & Society, 7 (4), 70-8. Honey, P., & Mumford, A. (992). The manual of Learnin Styles, Maidenhead: Peter Honey. IEEE (2002). IEEE Draft Standard for Learnin Object Metadata, IEEE P484.2./d6.4, Retrieved October 25, 2005, from, http://ltsc.ieee.or/w2/files/lom_484_2 v_final_draft.pdf. IMS. (200). IMS Global Learnin Consortium Inc., Learner Information Packae (LIP) Final Specification v.0, Retrieved October 25, 2005, from, http://www.imslobal.or/profiles/. Karampiperis, P., & Sampson, D. (2004). Adaptive Learnin Object Selection in Intellient Learnin Systems. Journal of Interactive Learnin Research, 5 (4), 389-409. Kay, J., & Holden, S. (2002). Automatic extraction of ontoloies from teachin document metadata. In Proc. of the ICCE 2002 Workshop on Concepts and Ontoloies in Web-based Educational Systems, Auckland, New Zealand, 25-28, Retrieved October 25, 2005, from, http://wwwis.win.tue.nl/~laroyo/icce2002_workshop/proc- Workshop-ICCE2002.pdf. 46

Khan, B. (200). Manain E-Learnin Strateies: Desin, Delivery, Implementation and Evaluation, Hershey, PA, USA: Idea Group Inc. Kravcik, M., & Specht, M. (2004). Flexible Naviation Support in the WINDS Learnin Environment for Architecture and Desin. Paper presented at the Third International Adaptive Hypermedia and Adaptive Webbased Systems Conference (AH2004), Auust 23-26, 2004, Eindhoven, The Netherlands. Meister, J. (2002). Pillars of e-learnin success, New York, USA: Corporate University Exchane. Paiva A., & Self J. (995). TAGUS - A User and Learner Modelin Workbench. International Journal of User Modelin and User-Adapted Interaction, 4 (3), 97-226. Papanikolaou, K., Grioriadou, M., Kornilakis, H., & Maoulas G. (2003) Personalisin the Interaction in a Web-based Educational Hypermedia System: the case of INSPIRE. International Journal of User Modelin and User-Adapted Interaction, 3 (3), 23-267. Sicilia, M. A., Garcia, E., Diaz, P., & Aedo, I. (2004). Usin links to describe imprecise relationships in educational contents. International Journal of Continuin Enineerin Education and Life-Lon Learnin (IJCEELL), 4 (3), 260 275. Simon, B., Dolo, P., Miklos, Z., Sintek, M., & Olmedilla, D. (2004). Conceptualisin Smart Spaces for Learnin. Journal of Interactive Media in Education, 9 (8), Retrieved October 25, 2005, from, http://www.l3s.de/~olmedilla/pub/jime04.pdf. Stewart C., Cristea A., Brailsford T., & Ashman H. (2005), Authorin once, Deliverin many: Creatin reusable Adaptive Courseware. Paper presented at the Web-Based Education Conference WBE'05, February 2-23, 2005, Grindelwald, Switzerland. Wilkie, D. (980). Pictorial Representation of Kendall s Rank Correlation Coefficient. Teachin Statistics, 2, 76-78. Wu, H., & De Bra, P. (200). Sufficient Conditions for Well-behaved Adaptive Hypermedia Systems. Lecture Notes In Computer Science, 298, 48-52. 47