2010 10th IEEE International Conference on Advanced Learning Technologies Ontology-Based Solution for Personalized recommendations in e-Learning Systems Methodological Aspects and Evaluation Criterias Mihaela brut. Florence sedes Institut de recherche en Informatique de Toulouse, Universite Paul Sabatier Toulouse. france Mihaela Brut, Florence. Sedes@irit.fr AbstractThe current paper expose a technique for developing background, and individual traits [4]. The profile could be systems adopting an ontology-based modeling of user profiles semantic networks (taxonomies, topic maps, or even and document models. Because the solution is situated at the ontologies)[3], [51 nterference of three domains (e-learning, semantic Web and In the el domain there are two main standards for adaptive hypermedia systems), the methodological aspects defining the user profile, where user's competences is the considered in developing such a solution are discussed with most important characteristic: the IMS Learner Information respect to the existing techniques in these domains. As well, Specification and the IEEE PAPI(Public And Private some evaluation criteria of such solution are discussed. while considering some existing systems that have similar characteristics to the proposed solution. In the Sw community, user competences were expressed through models such as HR-XML (a standard for exchanging Keywords: adaptive e-learning systems, recon data in the human relations domain ), or ResumeRDF tee ontology(defined for expressing curriculum vitae data via RDF constructs), or through XML/RDF/ OWL versions of . INTRODUCTION the ACM, ODP(Open Directory Project), or ECDL (European Computer Driving Licence)taxonomies 3 In the context where the number of available resources Moreover, some relations between concepts were defined in inside e-learning systems increases steeply, recommending order to refine the user profile: prerequisite, is-a, part-of[4] relevant resources is useful since it eliminates or reduces the The Recommender systems develop the user profile based time for browse and search, and also facilitates users to on the user navigation activity, considered in terms of items recognize what resources are interesting for them, since it is pages,(annotated) documents, etc. Data mining technologies often difficult for them to articulate their particular needs[l]. are applied to identify the current user activity in an expected Methods for improving the traditional recommendations sequence of tasks and to provide personalized task-level techniques are developed, such as combining multiple support[2] traditional techniques into a hybrid one, or integrating In the discussed context, the user profile should be based Semantic Web approaches in these techniques [2]. For be on e-learning standards, should express characteristics employed into an e-learning system, these techniques should required by AHS and expressed through an ontology and be accommodated with specific standards and structures 3]. should integrate the user activity The present paper presents the methodological aspects In [7] we presented a user profile developed on the top of for developing such a solution of personalized recom- the IEEE PAPI e-learning standard by extending its Learner mendations for e-learning systems, with respect to the Performance category. ACM topic hierarchy was chosen for existing techniques in the three interfered domains: e- expressing the user characteristics, grouped on three layers learning(EL), semantic Web(Sw) and adaptive hypermedia (overlapped on the user's knowledge, interests, and goals systems(AHS). The user profile( Section f), the document individual traits -see Section II) model (Section I), and the recommendation technique Competence. expressing the actual, alread Section IV) are discussed and some previously reported solutions are considered as illustrative examples. Section V Interests-the desired, foresighted, competencies presents a comparison of the proposed solution with the Fingerprints -the currently visited concepts via the existing systems and a discussion regarding an accurate annotations associated to the visited documents evaluation, The conclusions and research perspecti presented in the end of the pape *** IEEE P1484225- Draft Standard for Learning Technology USER PROFILE Public and Private In-formation(PAPI) for Learners(PAPI Learner Learner Performance Information. 2001 S adopt a feature-based modeling technique for 2http://www.hr-xml.org/,http://rdfs.org/resume-rdfy user profile, considering some important characteristics http://www.acm.org/about/class/,http://www.dmozorg/, the user as an individual: knowledge, interests, go http://www.ecdl.org 978-0-7695-4055-9/1052600◎2010IEEE DOI10.1109/CALT2010.136
Ontology-Based Solution for Personalized Recommendations in E-Learning Systems Methodological Aspects and Evaluation Criterias Mihaela Brut, Florence Sèdes Institut de Recherche en Informatique de Toulouse, Université Paul Sabatier Toulouse, France {Mihaela.Brut, Florence.Sedes}@irit.fr Abstract—The current paper expose a technique for developing a solution of personalized recommendations for e-learning systems adopting an ontology-based modeling of user profiles and document models. Because the solution is situated at the interference of three domains (e-learning, semantic Web and adaptive hypermedia systems), the methodological aspects considered in developing such a solution are discussed with respect to the existing techniques in these domains. As well, some evaluation criteria of such solution are discussed, while considering some existing systems that have similar characteristics to the proposed solution. Keywords: adaptive e-learning systems, recommendation techniques, ontology-based modeling I. INTRODUCTION In the context where the number of available resources inside e-learning systems increases steeply, recommending relevant resources is useful since it eliminates or reduces the time for browse and search, and also facilitates users to recognize what resources are interesting for them, since it is often difficult for them to articulate their particular needs [1]. Methods for improving the traditional recommendations techniques are developed, such as combining multiple traditional techniques into a hybrid one, or integrating Semantic Web approaches in these techniques [2]. For being employed into an e-learning system, these techniques should be accommodated with specific standards and structures [3]. The present paper presents the methodological aspects for developing such a solution of personalized recommendations for e-learning systems, with respect to the existing techniques in the three interfered domains: elearning (EL), semantic Web (SW) and adaptive hypermedia systems (AHS). The user profile (Section II), the document model (Section III), and the recommendation technique (Section IV) are discussed and some previously reported solutions are considered as illustrative examples. Section V presents a comparison of the proposed solution with the existing systems and a discussion regarding an accurate evaluation. The conclusions and research perspectives are presented in the end of the paper. II. USER PROFILE AHS adopt a feature-based modeling technique for the user profile, considering some important characteristics of the user as an individual: knowledge, interests, goals, background, and individual traits [4]. The profile could be represented based on a keywords set, on one or multiple semantic networks (taxonomies, topic maps, or even ontologies) [3], [5]. In the EL domain there are two main standards for defining the user profile, where user’s competences is the most important characteristic: the IMS Learner Information Specification and the IEEE PAPI (Public And Private Information)1 . In the SW community, user competences were expressed through models such as HR-XML (a standard for exchanging data in the human relations domain), or ResumeRDF ontology (defined for expressing curriculum vitae data via RDF constructs)2 , or through XML/RDF/ OWL versions of the ACM, ODP (Open Directory Project), or ECDL (European Computer Driving Licence) taxonomies 3 . Moreover, some relations between concepts were defined in order to refine the user profile: prerequisite, is-a, part-of [4]. The Recommender systems develop the user profile based on the user navigation activity, considered in terms of items, pages, (annotated) documents, etc. Data mining technologies are applied to identify the current user activity in an expected sequence of tasks and to provide personalized task-level support [2]. In the discussed context, the user profile should be based on e-learning standards, should express characteristics required by AHS and expressed through an ontology and should integrate the user activity. In [7] we presented a user profile developed on the top of the IEEE PAPI e-learning standard by extending its Learner Performance category. ACM topic hierarchy was chosen for expressing the user characteristics, grouped on three layers (overlapped on the user’s knowledge, interests, and goals individual traits - see Section II): • Competences – expressing the actual, already acquired, competences; • Interests – the desired, foresighted, competencies; • Fingerprints – the currently visited concepts via the annotations associated to the visited documents. 1 * * *, IEEE P1484.2.25 - Draft Standard for Learning Technology. Public and Private In-formation (PAPI) for Learners (PAPI Learner) — Learner Performance Information, 2001 2 http://www.hr-xml.org/, http://rdfs.org/resume-rdf/ 3 http://www.acm.org/about/class/, http://www.dmoz.org/, http://www.ecdl.org/ 2010 10th IEEE International Conference on Advanced Learning Technologies 978-0-7695-4055-9/10 $26.00 © 2010 IEEE DOI 10.1109/ICALT.2010.136 469
The first two layers are developed off-line through a already mentioned development of hybrid techniques rule-based technique, while the last layer is developed in real and the integration of semantic Web technologies could time, through the recommendation technique exposed in enhance the quality of recommendations Section iv Essentially for any recommendation technique is to analyze the user's navigational activity(by using IIL. DOCUMENT MODEL EL standards such as SCORM, IEEE LOM(Learning abstraction level approaches considered the concept-based Object Metadata)or ADL are conceived for learning navigation(where each concept and page is a navigation management purposes, and their main objective is to hub, as in the KBs Hyperbook system), document cluster facilitate the reuse of the Learning Objects(LOs) level navigation [9]or task-oriented navigation [2] Sw developed vocabularies such as RDF, DCMI, FOAF, The methodology for integrating ontologies into DOAP, SIOC, OpenGUID, as well as particular ontologies, recommenders involve three steps [13] sed to annotate certain information type, which thus gains a semantic meaning transparent to computers generating domain ontology an n el systems, ontologies could be used to exclusively Pattern discovery: to analyze user choices in order tate materials or in combination with e-learning to establish semantic usage patterns, standards [8]. Various relation types and even roles and/or Recommendations: to match user profile to domain weights associated to these relations [9 were adopted in order to refine the ontology-based annotations of the LOs In [14] we proposed a recommendation approach whose The annotation process is mostly manually or semi- novelty consists in supervising the user conceptual automatically accomplished. navigation through the ontology instead of his/her site In recommender systems, the documents are navigation for each visited document. its annotations are mostly as items as a whole, or as pages with certai considered in order to define user fingerprints through structures, and they are automatically processed in ontology concepts; as well, by adopting a measure of Among the current techniques for document annotation knowledge is integrated into recommendation algorithm main develop the document model [101 similarity between concepts, the ontology-based dor according to a domain model, the latent tic indexing The first two steps from the above-mentioned technique [11] could be mentioned,or knowledge methodology are facilitated: documents are considered as representation models and methods that are typical to already annotated based on an existing ontology, while the artificial intelligence domain(such as neural networks, user fingerprints are analyzed in rapport with his existing semantic networks, bayesian networks)[10] interests and competences(which play a pattern role In our context. in order to be conformed to the ex XIgencles Concretely, the recommendation technique is a hybrid of the three domains. a document model should be based on one that involves two phases [14] EL and Sw standards, expressed through ontological Collaborative filtering phase: user conceptual constructs, and automatically developed navigation is tracked in order to predict the next In [12] we presented a document model built on top of concept which will be focused by the user, according the IEEE/LOM e-learning standard by extending its to his fingerprints and interests profile; Classification category and by defining three relation types Content-based filtering: this concept is used in order between LOs and ontology concepts: isOn Topic. to select the documents to be effectively usesThe Concept, makesReference To recomme ded. in concordance with the user We also developed an automatically annotation technique competences profile that split the document in three classes for generating three For testing purposes, we considered a fragment of the semantic relations respectively: title and subtitles(headings), ACM topic hierarchy and we developed a training set of hyperlinks and bibliographical references, document body. fingerprint profiles values considering differentuser For each document class, a Latent Semantic Indexing categories(beginner, intermediate, advanced), in different technique is applied and enhanced by a Wordnet-based phases of course attendance. As well, we developed a set of documents annotations(according the technique exposed in Section Ill) and we used it as training set for the second recommender. a collection of 10 documents for each RECOMMENDATION TECHNIQUE ontology concept was used for testing purposes. The most Personalized access to the information takes a variety of dvanced users, while the pertinence decreases for the ccurate recommendations were encountered for the orms inside adaptive hypermedia systems: personalized intermediate and beginner recommendation techniques were developed: content-based could be the particularity of advanced users to be more ndations and collaborative filtering. As well, focused on a precise topic in their actions. V. EVALUATION POSSIBILITIES "http://www.we ning. com/primer standards. aspx Normally, a recommendation technique is evaluated with http://semanticweb.org/wiki/ontologies pect to the similar approaches by comparing the results
The first two layers are developed off-line, through a rule-based technique, while the last layer is developed in real time, through the recommendation technique exposed in Section IV. III. DOCUMENT MODEL EL standards such as SCORM, IEEE LOM (Learning Object Metadata) or ADL 4 are conceived for learning management purposes, and their main objective is to facilitate the reuse of the Learning Objects (LOs). SW developed vocabularies such as RDF, DCMI, FOAF, DOAP, SIOC, OpenGUID, as well as particular ontologies5 , used to annotate certain information type, which thus gains a semantic meaning transparent to computers. In EL systems, ontologies could be used to exclusively annotate materials or in combination with e-learning standards [8]. Various relation types and even roles and/or weights associated to these relations [9] were adopted in order to refine the ontology-based annotations of the LOs. The annotation process is mostly manually or semiautomatically accomplished. In recommender systems, the documents are treated mostly as items as a whole, or as pages with certain fixed structures, and they are automatically processed in order to develop the document model [10]. Among the current techniques for document annotation according to a domain model, the latent semantic indexing technique [11] could be mentioned, or some knowledge representation models and methods that are typical to artificial intelligence domain (such as neural networks, semantic networks, bayesian networks) [10]. In our context, in order to be conformed to the exigencies of the three domains, a document model should be based on EL and SW standards, expressed through ontological constructs, and automatically developed. In [12] we presented a document model built on top of the IEEE/LOM e-learning standard by extending its Classification category and by defining three relation types between LOs and ontology concepts: isOnTopic, usesTheConcept, makesReferenceTo. We also developed an automatically annotation technique that split the document in three classes for generating three semantic relations respectively: title and subtitles (headings), hyperlinks and bibliographical references, document body. For each document class, a Latent Semantic Indexing technique is applied and enhanced by a Wordnet-based approach. IV. RECOMMENDATION TECHNIQUE Personalized access to the information takes a variety of forms inside adaptive hypermedia systems: personalized search, focused crawling, recommenders. Two main recommendation techniques were developed: content-based recommendations and collaborative filtering. As well, as 4 http://www.webbasedtraining.com/primer_standards.aspx 5 http://semanticweb.org/wiki/Ontologies already mentioned, the development of hybrid techniques and the integration of semantic Web technologies could enhance the quality of recommendations. Essentially for any recommendation technique is to analyze the user’s navigational activity (by using some consecrated data mining algorithms). Some “higher” abstraction level approaches considered the concept-based navigation (where each concept and page is a navigation hub, as in the KBS Hyperbook system), document cluster level navigation [9] or task-oriented navigation [2]. The methodology for integrating ontologies into recommenders involve three steps [13]: • Data preparation: to analyze documents for generating domain ontology; • Pattern discovery: to analyze user choices in order to establish semantic usage patterns; • Recommendations: to match user profile to domain model. In [14] we proposed a recommendation approach whose novelty consists in supervising the user conceptual navigation through the ontology instead of his/her site navigation: for each visited document, its annotations are considered in order to define user fingerprints through ontology concepts; as well, by adopting a measure of similarity between concepts, the ontology-based domain knowledge is integrated into recommendation algorithm. The first two steps from the above-mentioned methodology are facilitated: documents are considered as already annotated based on an existing ontology, while the user fingerprints are analyzed in rapport with his existing interests and competences (which play a pattern role). Concretely, the recommendation technique is a hybrid one that involves two phases [14]: • Collaborative filtering phase: user conceptual navigation is tracked in order to predict the next concept which will be focused by the user, according to his fingerprints and interests profile; • Content-based filtering: this concept is used in order to select the documents to be effectively recommended, in concordance with the user competences profile. For testing purposes, we considered a fragment of the ACM topic hierarchy and we developed a training set of fingerprint profiles values considering different user categories (beginner, intermediate, advanced), in different phases of course attendance. As well, we developed a set of documents annotations (according the technique exposed in Section III) and we used it as training set for the second recommender. A collection of 10 documents for each ontology concept was used for testing purposes. The most accurate recommendations were encountered for the advanced users, while the pertinence decreases for the intermediate and beginner users. A possible explanation could be the particularity of advanced users to be more focused on a precise topic in their actions. V. EVALUATION POSSIBILITIES Normally, a recommendation technique is evaluated with respect to the similar approaches by comparing the results 470
obtained over the same data sets. Because a completely integrating them with ontology-based knowledge similar approach doesnt exist(which to employ a similar while benefitting by a stable set of user profiles. That modeling and recommendation technique for the case of e- is what we intend to do further learning systems), nor the corresponding data sets, an recommendation technique, and we hope to find in meantime evaluation of algorithmic performance is not possible for the some public datasets available, provided by the existing proposed recommendation technique n our context, the evaluation could consist into a discussion considering some comparable aspects such as REFERENCES Specific type of AHs that adopted a domain ontology for developing the user and document []RFarzan, PBI models, regardless it concerns or not the el domain Recommendation System", in V.P. Wade, H. Ashman, B. Smyth Considered user traits Eds ) Proceedings of AH 2006, Ireland, LNCS 4018, Springer 2006. Ontology adopted for document modeling [2]X Jin, Y. Zhou, B Mobasher, " Task-Oriented Web User Modeling Algorithmic solution employed for developing Recommendation, In Proceedings UMOS, Edinburgh, adaptive support July 2005, LNAI 3538, pp. 109-118, Springer, 2005 Data set used for test 3] P. Dolog, M. Schafer, "Learner Modeling on the Semantic Web.In oc. of Pers Web05, Workshop on Personalization on the Semantic In order to limit this comparison, we selected some AHS Web at 1Oth International User Modeling Conference, 2005 systems that employ similar user traits as RecOnto(the [4] P Brusilovsky, E Millan, "User Models for Adaptive Hypermedia recommendation technique described above): competences and Adaptive Educational Systems", in P. Brusilovsky, A. Kobsa, W (C), interests n) or fingerprints(F). Table I presents a Nejdl(Eds ) The Adaptive Web, LNCS 4321, Springer, 2007 nthesis of this comparison 5]J. Kay, A. Lum, "Ontologies for Scrutable Learner Modeling Adaptive E-Learning", In: Aroyo, L, Tasso, C.(eds )Proc. of TABLE I COMPARING RECONTO TO EXISTING SYSTEMS AH2004 Workshops Eindhoven, 2004, Pp. 292-301 Domain, Task, and User Models for System User Docum. Adap Data 6 P. Brusilovsky, D w. Coo Leake, DB(eds)Proc. of IUI 2001, ACM Press, 2002, pp. 23.10. earning Bayes [ M. Brut, L. Asandului, G. Grigoras, "A Rule-Based Approach for networks Systems", in M. Perry, H. Sasaki, M. Ehmann, G. Ortiz Bellot(Eds ) Proceedings of ICIW 2009, IEEE Computer Society, Venice, 2009 based between [8]HS. Al-Khalifa D. Hugh,"The Evol Standards to Semantics in E-Learning Applications", Proceedings of Hypertext 2006, ACM Press, 2006 my of job recom- 19]N. Henze, w. Nejdl, ""Adaptation in open corpus hypermedia", Int J. of Artificial Intelligence in Education 12, 4, 2001, pp. 325-350 [10A. Micarelli, F. Sciarrone, M. Marinilli, M,"Web document 6 Neidl mender Adaptive Web: Methods and Strategies of Web Personalization, Persona F ODP Lecture Notes in Computer Science, Vol 4321. Springer, 2007 coloring [11] B. Sarwar, G. Karypis, J.A. Konstan, J. Riedl, "Incremental SVD- RecOnto C+l+ ontology Hybrid kNN Own Proceedings of CIT 2002, IEEE CS Press, 2002 + roles NN [12]M. Brut, F. Sedes, T, Jucan, R. Grigoras, V. Charvillat,"An VI. CONCLUSION AND FURTHER WORK EDL2 L Conference,wc200 Milano,al, Springer,28么° The paper presented some general methodological [3] H. Dal, B. Mobasher,"A Road Map to More Ettective Web ects to which are constrained the ahs systems that adopt ontology-based modeling in the context of EL domain. A H.R. Arabnia, Y. Mun(Eds ) Proceedings ternational Conference on Internet Computing IC 2003, Volume I concrete solution was also presented, which respects these CSREA Press 2003, pp 58-64 aspects. However, these aspects apropriated into [14] M. Bn variety of ways by other AHS systems, and thus it is E-Learning Recommender Systems", in Chevalier, M, Julien, C impossible to establish some precise evaluation criteria. ule- Dupuy, C, Collaborative and Social Information Retrieval and moreover, the results obtained by each particular system are Access: Techniques for Improved User Modeling, IGI Global, 2008 often dependent of the user types(beginner, normal, expert) [15] P. Haase, N. Stoja J. Volker. Y. Sure. "Personalized Information Retrieval in Bibster, a Semantics-Based Bibliographic employed for testing purposes. Thus, a coherent evaluation Peer-to-Peer System", Proc. of l-KNOW 05, Austria, 2005 for an AHs adopting ontology-based modeling and conceived for a particular domain such as EL should be [16]SE. Middleton,NRShadbolt,D.CDe Roure,D.C,"Ontological User profiling in R ender Systems", ACM Transactions on focused on a particular adaptive functionality (such as Information Systems 2, No. 1, January 2004, pp. 54-88 recommendations) and should test various algorithmic [17) F. Tanudjaja, L. Mui, "Persona: A Contextualized and Personalized recommendation olutions possibili Web Search". In: Proceedings of HICSS 2002, Hawai, IEEE Computer Society Press 2002
obtained over the same data sets. Because a completely similar approach doesn’t exist (which to employ a similar modeling and recommendation technique for the case of elearning systems), nor the corresponding data sets, an evaluation of algorithmic performance is not possible for the proposed recommendation technique. In our context, the evaluation could consist into a discussion considering some comparable aspects such as: • Specific type of AHS that adopted a domain ontology for developing the user and document models, regardless it concerns or not the EL domain; • Considered user traits; • Ontology adopted for document modeling; • Algorithmic solution employed for developing adaptive support; • Data set used for test. In order to limit this comparison, we selected some AHS systems that employ similar user traits as RecOnto (the recommendation technique described above): competences (C), interests (I) or fingerprints (F). Table 1 presents a synthesis of this comparison. TABLE I. COMPARING RECONTO TO EXISTING SYSTEMS System User profile Docum. model Adaptation support Algorithmic solution Data set KBS [9] C+I concept network learning path Bayes networks Own developed Bibster [15] I ACM taxonomy querybased retrieval Similarity between concepts Community dev. Course Agent [13] I Taxonomy of job profiles Course recommender Case-based algorithm Community dev. Foxtrot [16] F ACM Hybrid recommender IBK + AdaBoost Own dev. Persona [17] F ODP querybased retrieval tree coloring method Own dev. RecOnto C+I+ F ontology + roles Hybrid Recomm ender kNN + kNN Own dev. VI. CONCLUSION AND FURTHER WORK The paper presented some general methodological aspects to which are constrained the AHS systems that adopt ontology-based modeling in the context of EL domain. A concrete solution was also presented, which respects these aspects. However, these aspects are appropriated into a variety of ways by other AHS systems, and thus it is impossible to establish some precise evaluation criteria. Moreover, the results obtained by each particular system are often dependent of the user types (beginner, normal, expert) employed for testing purposes. Thus, a coherent evaluation for an AHS adopting ontology-based modeling and conceived for a particular domain such as EL should be focused on a particular adaptive functionality (such as recommendations) and should test various algorithmic recommendation solutions, as well as possibilities of integrating them with ontology-based domain knowledge, while benefitting by a stable set of user activity profiles. That is what we intend to do further with the RecOnto recommendation technique, and we hope to find in meantime some public datasets available, provided by the existing systems. REFERENCES [1] R. Farzan, P. Brusilovsky, “Social Navigation Support in a Course Recommendation System”, in V.P. Wade, H. Ashman, B. Smyth (Eds.), Proceedings of AH 2006, Ireland, LNCS 4018, Springer 2006. [2] X. Jin, Y. Zhou, B. Mobasher, “Task-Oriented Web User Modeling for Recommendation”, In Proceedings UM'05, Edinburgh, Scotland, July 2005, LNAI 3538, pp.109-118, Springer, 2005. [3] P. Dolog, M. Schäfer, “Learner Modeling on the Semantic Web”. In: Proc. of PerSWeb’05, Workshop on Personalization on the Semantic Web at 10th International User Modeling Conference, 2005. [4] P. Brusilovsky, E. Millán, “User Models for Adaptive Hypermedia and Adaptive Educational Systems”, in P. Brusilovsky, A. Kobsa, W. Nejdl (Eds.), The Adaptive Web, LNCS 4321, Springer, 2007. [5] J. Kay, A. Lum, “Ontologies for Scrutable Learner Modeling in Adaptive E-Learning”, In: Aroyo, L., Tasso, C. (eds.) Proc. of AH'2004 Workshops. Eindhoven, 2004, pp. 292-301. [6] P. Brusilovsky, D.W. Cooper, “Domain, Task, and User Models for an Adaptive Hypermedia Performance Support System”, in Gil, Y., Leake, D.B. (eds.) Proc. of IUI 2001, ACM Press, 2002, pp. 23-30. [7] M. Brut, L. Asandului, G. Grigoras, “A Rule-Based Approach for Developing a Competency-Oriented User Model for E-Learning Systems”, in M. Perry, H. Sasaki, M. Ehmann, G. Ortiz Bellot (Eds.), Proceedings of ICIW 2009, IEEE Computer Society, Venice, 2009. [8] H.S. Al-Khalifa D. Hugh, “The Evolution of Metadata from Standards to Semantics in E-Learning Applications”, Proceedings of Hypertext 2006, ACM Press, 2006 [9] N. Henze, W. Nejdl, “Adaptation in open corpus hypermedia”, Int. J. of Artificial Intelligence in Education 12, 4, 2001, pp. 325-350 [10] A. Micarelli, F. Sciarrone, M. Marinilli, M., “Web document modeling”, in P. Brusilovsky, A. Kobsa, W. Nejdl (eds.), The Adaptive Web: Methods and Strategies of Web Personalization, Lecture Notes in Computer Science, Vol. 4321. Springer, 2007 [11] B. Sarwar, G. Karypis, J.A. Konstan, J. Riedl, “Incremental SVDBased Algorithms for Highly Scaleable Recommender Systems”. Proceedings of CIT 2002, IEEE CS Press, 2002 [12] M. Brut, F. Sèdes, T. Jucan, R. Grigoras, V. Charvillat, “An Ontology-Based Modeling Approach for Developing a Competencies-Oriented Collective Intelligence”, Proceedings of the ED-L2L Conference, WCC 2008, Milano, Italy, Springer, 2008 [13] H. Dai, B. Mobasher, “A Road Map to More Effective Web Personalization: Integrating Domain Ontologies with Web Usage Mining”, in H.R. Arabnia, Y. Mun (Eds.), Proceedings of International Conference on Internet Computing IC 2003, Volume 1. CSREA Press 2003, pp. 58-64. [14] M. Brut, F. Sèdes, C. Zayani, “Adopting Ontology-Based Modeling in E-Learning Recommender Systems”, in Chevalier, M., Julien, C., Soulé-Dupuy, C., Collaborative and Social Information Retrieval and Access: Techniques for Improved User Modeling, IGI Global, 2008 [15] P. Haase, N. Stojanovic, J. Volker, Y. Sure, “Personalized Information Retrieval in Bibster, a Semantics-Based Bibliographic Peer-to-Peer System”, Proc. of I-KNOW ’05, Austria, 2005. [16] S.E. Middleton, N.R. Shadbolt, D.C. De Roure, D.C., “Ontological User profiling in Recommender Systems”, ACM Transactions on Information Systems, Vol. 22, No. 1, January 2004, pp. 54-88 [17] F. Tanudjaja, L. Mui, “Persona: A Contextualized and Personalized Web Search”. In: Proceedings of HICSS 2002, Hawai, IEEE Computer Society Press 2002. 471