You are on page 1of 6

Personalization on the Web of Data and New Paradigms for Distributed and Open User Modeling

Lora Aroyo
VU University Amsterdam FEW-CS, De Boelelaan 1081 1081 HV Amsterdam, the Netherlands +31 20 5982868

Geert-Jan Houben
TU Delft EWI/ST/WIS, P.O. Box 5031 2600 GA Delft, the Netherlands +31 15 2787486

l.m.aroyo@cs.vu.nl ABSTRACT
With more personal data being out there on the Web of Data and with an increasing level of personalization in the way data is offered on the Web, the traditional role of user modeling is changing. Traditionally, user modeling was closely aligned with the user-adaptation offered by specific adaptive systems, and the research field of User Modeling reflected this in its assumptions. For the Web of Data the situation has changed, user modeling has become both massively distributed and open, and new assumptions are needed. This has consequences both for the fields of User Modeling and Adaptive Systems, in terms of their assumptions and research questions, but also for Web Science as the role of user model knowledge changes the way the web is used. Therefore, the many new research challenges in engineering and studying distributed and open user modeling have to be approached in the triangle between User Modeling, Adaptive Web-based Systems, and Web Science.

g.j.p.m.houben@tudelft.nl
Just like the Web of Data itself is a consequence of a desire to move towards a more distributed and open environment of data, also for user modeling and for user model knowledge we need to acknowledge that distribution and openness are necessary ingredients of the new Web. Personalization on the Web of Data expects the data to be part of a distributed and open environment, and now also needs to do the same for the user model knowledge that represents the other side of the coin when it comes to useradaptation and personalization. It is not difficult to see from current research in user modeling that significant advances of adaptation and personalization on the Web can only be realized with truly distributed and open user modeling approaches. To obtain effective approaches for distributed and open user modeling, a combined effort is needed from three research fields. The fields of Adaptive Web-based Systems and User Modeling obviously need to continue to interact to define the assumptions and research questions that relate to the use of user model knowledge for adaptation, but more than ever also the consequences of distributed and open user modeling on the Web have to be included in the considerations. Web Science efforts in this direction include the engineering challenges to deal with distributed and open user model knowledge on a true Web scale, but also the study of the many issues (social, legal, ethical, behavioral, etc.) that follow from a different way of user modeling on the Web. Where for example semantics appear to offer great help for connecting the data on the Web of Data, in order to guide the user modeling research in the proper effective direction, for example using semantic techniques in user model representation, in the triangle between User Modeling, Adaptive Web-based Systems and Web Science research needs to be connected and targeted to have a chance to meet the research challenges. In this paper we take a look back at user modeling and adaptation in section 2. We then consider the effects of (semantic) linking data, adaptation and users as the state is now in section 3. In section 4 we then consider what is ahead and lay out the challenges that await the research fields, before concluding in section 5.

Keywords
User modeling, adaptive web-based systems, web science, linked open data, adaptation, personalization, recommendation, distribution, open.

1. INTRODUCTION
In this paper we reflect on the changes that are awaiting the field of User Modeling when it tries to meet the demands of personalization on the Web of Data. With the new nature of the Web, the desire to make the Web personalized in its offerings to the large variety of its users asks for new and different solutions. In traditional Web-based systems the role of the adaptation was heavily determined by the constraints and assumptions of the system, and the user modeling that would support this assumption would operate from the same constraints and assumptions. More recently, we have already seen research efforts leading to small steps in the direction of collaboration or integration between adaptive systems, but in order to obtain true personalization on the Web of Data a more drastic change is needed.

2. USER MODELING
In the evolution of the World Wide Web, the Web of documents, it became quickly apparent that although the hypertext basis for the Web was a key factor in its success, a single hyperlink structure of nodes and links would not serve all the people using it equally well. As a consequence of the goal to adapt the offerings to the user, the field of adaptive hypermedia started exploring how the hyperlink basis communicating content to the users could be

made adaptive to each separate user. At first, this was approached in single closed hyperdocuments and in closed hypermedia systems, then in open hypermedia systems, and later as part of systems operating on the Web. The hyperlink structure that best communicates content to a user depends on the users goal, the users context or the users background, to name just a few of the aspects that influence the users reception of hyperlinked documents. An easy example can be drawn from the field of education, where a teacher designing a hyperlink structure between documents in order to teach a student a given subject will consider a number of aspects that determine the student to decide which pages to create and which links to create between those pages. An important element in these considerations of the teacher will be the knowledge that the student acquires while reading the pages and following the links. Obviously, the teacher will try to apply known pedagogic principles to make the student go through the content in accordance with the knowledge accumulated on the subject and will therefore create the pages and links to help in that process of carefully planned knowledge accumulation. It is important to note that for each link and each page the teacher designs in that structure, it is essential that the teacher makes an assumption or pre-condition of the knowledge the student will have when being faced with that page or link: this knowledge is not just knowledge acquired before the start of the course, but also the knowledge obtained while interacting with the content and even the behavior in that learning process can be relevant for the teachers decisions. Because of the easy-to-understand aspects of adaptation in educational applications [2], educational examples have traditionally played an important role in the field of adaptive hypermedia, but also examples from domains such as tourism and e-commerce haven drawn a lot of attention from researchers. The research field of adaptive hypermedia [4] has since long been the central area for research in making content and navigation adaptive to the user, and thus turning hypermedia content into adaptive hypermedia content. This research includes the creation of systems to provide users with adaptive access to hypermedia-based content and the analysis of the usage of such systems and applications. From the start of the area, there has been a lot of attention to the creation of systems with a restricted and well-defined scope (in terms of the hypermedia structure): by locally managing and knowing the content and its structure and by making assumptions about the users knowledge and interaction for that content, it becomes possible to specify exactly what the adaptive system has to do in the interaction with the user. In an evolution of this work where adaptive hypermedia systems met the Web [5,6], we have seen a new generation of adaptive Web-based systems, where the scoping has been relaxed. As part of this evolution of adaptive functionality, a lot of research has in the last two decades been targeting the tooling for adaptive hypermedia and adaptive web-based systems. In [16] and its preceding work we have witnessed the consistent extension of the possibilities to specify and execute user-adaptation in hypermedia settings. As part of this trend, we can observe how not only the (expressive) power of the adaptation engines and mechanisms has been extended, but also how the integration of content from outside the original scope is supported through integration and linking. However, next to creating adaptive solutions and support, understanding the usage of adaptive hypermedia is also essential to obtain good user-adaptive hypermedia solutions. Without a solid understanding of how users experience the adaptation,

adaptive hypermedia cannot effectively be applied as an instrument in relevant situations. This brings up the connection to the research area of user modeling, e.g., [7], where the central focus has been for decades to construct theory and techniques for gathering and representing user knowledge. For any system that tries to acknowledge the users state (of knowledge, context, background etc.), whether adaptive hypermedia system or any other intelligent solution, it is crucial that the system knows who the user is at that moment. User modeling is in this light perceived as the techniques, often AI-based techniques, that determine what are relevant attributes of the user for effective adaptation and how to obtain good values for those attributes. In domains such as recommendation and teaching user modeling has advanced enormously and through establishing relevant theories about the connection between users and adaptation it has become possible to make relevant user knowledge available for adaptive applications. The outcome of the user modeling effort is a user model, and the link between the fields of user modeling and adaptive hypermedia [13] is that in the architecture and setup of adaptive hypermedia systems, a user model plays a crucial role. Until now we have mainly discussed the possibility to adapt, but this adaptation is defined on top of a user model. In most adaptive hypermedia solutions, exemplified by the standard reference model from [9], this user model exists in an explicit way, as a model of user knowledge that is explicitly available for the system to base its adaptation decisions on. So, after knowing the user the system can then concentrate on figuring out how to present content and navigation for the known user. Following [9], the general architecture discerns a domain model to describe the content of the application, a user model to describe the user, and an adaptation model that describes how domain and user knowledge are both used to decide about presenting pages and links.

Figure 1. General Adaptation Architecture

The user model is, as said, an explicit piece of data that represents the relevant aspects of the user, relevant for adaptation. Since in many applications the relevant aspects of the user include the users knowledge about the content that is presented in the application, the systems use most often an overlay model, which means that the user model is based on the domain model, for example by associating with each concept in the domain model a value representing the users knowledge about that concept. The type of knowledge that is typically contained in the user model is of course depending on the goal and purpose of the application. The aspects that we see most frequently, although not all at the same time, i.e. not all in the same application, are: history, background, preferences, level of knowledge, goals and tasks,

context of work, meta-cognitive skills, personality traits, affective states, attitudes, etc.

3. LINKING DATA AND USERS 3.1 Web of (Linked) Data


In order to understand the evolution in adaptation, we first observe the evolution in the Web. From a World-Wide Web of documents, we have now a moved to a Web of Data. Following the advances in the Semantic Web and the opportunities offered in the cloud of Linked Open Data [15], the basis for the content we concentrate on in our applications has changed. So, the basis for adaptation has changed. It therefore became straightforward to investigate how this evolution towards the Web of Data affected adaptation.

Figure 3. Linking of Users

3.2 Linking Data for Adaptation


In Section 2, we have given a background in user-adaptive systems. When the original adaptive hypermedia systems started their evolution towards adaptive Web-based systems, the main advantage with the role of the Web was the interoperability between applications. It will be obvious that in order to obtain good adaptive applications, a major investment is needed in terms of eliciting the relevant knowledge on the domain content, the users, and the effects of adaptation. The very first aspect where interoperability could play a beneficial role is in sharing and reusing of domain content (in the domain model following [9]).

First of all, inside specific domains such as education it is easy to see the opportunities for benefits when for users that use a similar application knowledge is shared between domain-specific systems and applications [11]. If in this example knowledge about students can be shared and exploited between learning management systems over the boundaries of different courses on the same or related topics, then obviously richer, more and more relevant, student knowledge can be elicited. Second, with many Web 2.0 and social networking applications entering the arena, the sharing of persons and personal data at a Web scale became a topic. Here we can discern two aspects: the identification of people, i.e. users, and the description of peoples attributes, i.e. user properties.

3.3.1 User Identification


For the identification of users across applications, we can rely on the many initiatives that also serve the wave of social networking applications in Web 2.0. On the one hand, users themselves can use identity-based protocols as OpenID to link their various identities on the Web. On the other hand systems often use authentication mechanisms, e.g. basic http authentication or open protocols for secure API authorization, like OAuth.

3.3.2 User Property Representation


Figure 2. Linking of Data For the representation of user properties the field is still wide open. To understand the main issues and difficulties here, we need to know that most user models are overlay models in one way or another: it means that on top of a domain model it is specified what is known about the relation between the user and a domain element. To make it concrete, we can consider an educational setting where the user model is used to capture a students knowledge on the concepts in a given domain. Then, the user model will contain knowledge that allows getting an indication of the knowledge level for each domain concept. It will be clear that the mapping and alignment of the domain concepts across different applications, when sharing such user models, reduces to the normal mapping and alignment of knowledge structures: if the educational application P about Programming likes to use student knowledge from an application J about Java, then obviously in the design of P the relation between concepts from J and concepts from P will be specified. More specific however for user models is the problem how to share, exchange and interpret the knowledge levels: in the example of P and J, it means that P needs to know how J represents what the students knows about a given Java concept, before it can decide what this means for its own representation; if J associates one of the values passed and not-passed with each concept and P uses a percentage to indicate how well a student has learned a concept, then in the design of P it has to be decided what <C ; passed> means in

With applications that exploit external data from the Web, the originally closed systems open up their content. This was first started off in the adaptive hypermedia area with work in open hypermedia systems [3], separating the linking from the content, and later it became mainstream with content being retrieved from the Web, often based on metadata in Semantic Web-enhanced scenarios. It is not difficult to see how the integration of data in the spirit of the Semantic Web and Linked Data [15] can benefit the adaptive applications with better content: in that sense these adaptive applications do not build an exception, and they can use the many assets that are there to map and align content (in or across domains).

3.3 Linking Users for Adaptation


Next to the semantic interoperability for the content that is used in the applications, on the Web it became possible to start sharing knowledge about the users.

terms of a percentage p for a concept D in J. Even from this simple example we can conclude that here in the mapping and alignment we have to deal with both (a) the knowledge about the domain and (b) the knowledge about the knowledge of that domain: from these the former can be perhaps easily solved in a particular application domain, but the latter depends heavily on the specific adaptive systems and their approach to user model representation. We do stress that in many adaptive systems, the user modeling is done in a specific way, leading to a great variety of solutions, not always based on overlay modeling, but overall the rather specific and often proprietary way of representing the user properties defines an integration problem. This lack of interoperability in representing user properties is one of the biggest hurdles in the integration of adaptive functionality and personalization at a Web scale.

specialized components, for example to accommodate the users preferences for adjusting presentations to certain display properties. The former, sharing user model knowledge, appears to be a problem that can reasonably be solved with experience from the Semantic Web, as for example [1,12,14] demonstrate. This is one leading scenario for sharing user model knowledge that we address further on in this paper. Motives for this sharing of user model knowledge can be found in the fact that cooperation between applications prevent that an applications perception of the user does not need to be built up entirely by the application, but can be constructed reusing knowledge that other applications have already gathered: for example, the risk of a cold start can be reduced, and by having more evidence available the possibility to derive better perceptions increase. However, besides the benefit for cooperating applications, there is another starting point that brings us to a similar scenario. That is when we take the perspective of the individual user that likes to manage herself the personal information that applications on the Web use to adapt the interaction to her. Specially, in the space of social networking sites we have seen first proposals, e.g. [10], that allow users to control themselves their profiles with multiple applications. For both angles to the same scenario, it is necessary to have ways to share and integrate user model knowledge. Two aspects that are essential in that respect are distribution and openness. The process of sharing and integrating user model knowledge can thus best be described as Distributed and Open User Modeling.

4. LINKING ADAPTIVE KNOWLEDGE


In the previous section, we have looked at part of the state of affairs in user model-based adaptation in the evolving Web. In this section we consider more closely the ambition of integrating and sharing adaptive and personalization functionality and knowledge and the requirements and challenges to advance with that ambition and do so at Web scale. This will help us define the evolving challenges for user modeling on the Web. First, we lay out the ambitions and challenges in terms of adaptation on the Web.

4.1 Adaptation on the Web of Data


On the Web we find many applications that provide adaptive access and querying to data on the Web. These applications include sources in many domains and what they have in common is that the user can interact with them and will be served individually. Examples can be found in websites that sell books or music, in applications that recommend TV-programs or travels, in applications that support education, etc. These applications typically work separately, and it is left to a particular application to decide to set up a managed interaction with other applications to improve and extend its adaptation capabilities by sharing with other applications. This kind of managed integration from the side of the application owner and provider is a standard integration problem that suffers from two main obstacles: Sharing user model knowledge; Sharing adaptation functionality. The latter, sharing adaptation functionality, is the most complex one, mainly because of the proprietary nature of many of the adaptation implementation mechanisms being used. In general, adaptation is realized by a diversity of solutions, and even when we concentrate on adaptive hypermedia-based solutions that share more or less the same approach, even then the reuse of adaptation from one application into the other one remains complex. The Grapple project [11] is making first steps in that direction. Motives for sharing adaptation can be found in the aim to choose the best possible adaptation for a user. When we again take the example of the educational application, we could imagine how finding a good and effective adaptation strategy, as a reflection of an effective learning strategy, is worth being reused by other applications, that then do not have to go through a longer process of determining an optimal strategy. Another motive could come from the delegation of certain adaptation functionality to

4.2 Distributed and Open User Modeling


In the scenario addressed above, we are confronted with distributed user model knowledge for which some kind of combination or integration is asked for. In this section we will reflect on a number of the research questions and challenges that need to be addressed to make a significant step forward in distributed and open user modeling. For this purpose we dissect the integration problem into its different elements: Identification of the user; Alignment of the concepts in the user models; Alignment of the user knowledge about the concepts in the user models.

4.2.1 User Identification


A first problem to solve is the identification of users: How do we identify a person (or an appearance of a person)? While this question obviously also was asked in the first generations of adaptive systems, it will be quickly clear that the assumptions then and now have changed, and that many more issues come into a play. Let us consider, the question from two perspectives. First, the perspective of a person: How can a person identify himself to an application? How can a person manage his identities (across multiple applications)?

Second, let us take the perspective of adaptive applications:

How can applications find a user (identity) in other applications (e.g. for the purpose of integrating and offering a better service)?

standard vocabulary for these properties is not an impossible task. So, this leads to the following questions: Which are the properties that are typically found in user model knowledge? How can we represent these user properties?

Trying to answer these questions brings up issues of trust and privacy, but also issues from social and behavioral sciences (when it comes to how people will or will not want to share or reveal their identities or data), and of course there are legal and political consequences. Where adaptation providers might have the best intentions for the users, the way in which they approach this will be crucial for the success of personalization. It is clear that the mostly technical solutions that are being proposed in the fields of User Modeling and Adaptive Web-based Systems are insufficient in their study of the complete problem. Standards for representing user identities help, as do tools to manage and map identities, but the biggest challenge is still wide open and that is related to all the non-technical aspects that determine identification. Distributed and open user identification is a first true Web Science topic that needs to be put strongly on the research agenda.

Of course, a start from the common core of properties that are known from current systems does not imply that on the open Web things could not be different, e.g. some user properties only show up on the open Web, but with a common core it becomes possible to study those effects. So, after the initial knowledge representation challenge, Web Science can kick in to study the effects at the scale of the Web. Note that using a standardized representation does not say that all applications access and use one and the same user profile. Allowing to have a shared vocabulary should not be mixed with having one profile for all applications. However, with the shared vocabulary and the possibilities to align user model knowledge, besides the technical issues, there is the obvious need to study the implications regarding trust, privacy, etc., to be able to find a way of working that satisfies all parties that (can, may and do) decide to connect. User knowledge alignment is a second true Web Science topic that needs to be put strongly on the research agenda.

4.2.2 Concept Alignment in User Models


After having established user identification, sharing and exchanging user model knowledge starts with concept alignment: How is conceptual knowledge integrated and shared? It is not difficult to see how this challenge reduces to standard concept mapping and alignment if we just look at the techniques. From a technical point of view, there is nothing special, i.e. which is not already studied in Web of Data research. Of course, for the purpose of subsequently being able to foster more sharing of user model knowledge and therefore provide better adaptive services, obviously any attempt to do more reuse and sharing in data and knowledge, e.g. using SKOS, RDF/OWL etc., is beneficial. So, here we do not elicit any specific challenges, but we do observe how a semantic-based integration of data (concepts) can improve the chances for sharing user model knowledge about the data.

4.2.4 Openness and scrutability


Most of the considerations in the previous subsections relate to the distribution of users and user knowledge. Closely connected is the openness that comes with the Web of Data. Where this Web has evolved from the exploitation of open data, one can now consider the question how open user knowledge can be: How can user knowledge be made open to allow systems and applications to use it for personalization? This main question brings up many challenging problems, for which researchers need to find the answers. We stress that these answers are not the technical solutions that show how user knowledge could be published openly if we would forget about all legal, ethical etc. issues: it is exactly the non-technical part of the problem that should be leading in the study of how openness in user modeling can benefit personalization. Here, new paradigms are needed to understand how users and personalization providers perceive their mutual benefits. These new benefits might then require new technical solutions, but at first sight it appears that Web Science first needs to address how the publication of user knowledge can serve the purpose of the user. One aspect of open publication relates to a fundamental element of user modeling: scrutability. There should be a close interaction between the user and the system that maintains user knowledge about the user. Next to the control the user has to have of the user knowledge as it is shared or not shared between applications, it is also considered important to what level the user is able to inspect and alter her own user model knowledge. Obviously, there will be open and proprietary (to the provider) user knowledge, but system-specific mechanisms to elicit and verify user knowledge, e.g. [8], are to be extended for use in the context of the distributed and open Web. The openness of user knowledge is a third true Web Science topic that needs to be put strongly on the research agenda.

4.2.3 User Knowledge Alignment in User Models


With users identified and data aligned, the third major step is the alignment of user knowledge, i.e. the relation between the user and the data. The main question here is: How can user model knowledge be shared between adaptive applications? The most promising answer to this question implies that user model knowledge, even when gathered through very diverse techniques, should have a standard semantic representation. In that representation, e.g. through annotation, we then include for all user model knowledge the necessary provenance information about the user model knowledge. It is that (user model-specific) provenance information that helps to communicate what the user model knowledge represents. Then, this representation can be used to exchange between applications. Technically, this would not be difficult, but determining the relevant provenance information is a wide open issue. Obviously, there will be no standard representation that fits with all of the current tools and applications, but if we look at user modeling theories and experience we can see that the properties that are usually included (in studies in closed applications and systems) are more or less known for most cases, and to create a

5. CONCLUSION
In this paper we have taken a look at the current state of affairs in user modeling, and from the challenges that personalization on the Web of Data implies, we have derived a number of research questions that should be high on the research agenda. The three main subjects for further research are distributed and open user identification, user knowledge alignment, and openness of user knowledge. At the same time, we have shown how the traditional collaboration between the fields of User Modeling and Adaptive Web-based Systems could and should collaborate more intensively with Web Science. In that triangle it becomes possible to exploit the large number of theories and results for user modeling and adaptation on the new Web and study how well they perform, while also providing a good basis for investigating new paradigms for distributed and open user modeling.

[6] Brusilovsky, P., Maybury, M.T. From adaptive hypermedia to the adaptive web. Commun. ACM 45 (5) (2002) 3033. [7] Conati, C., McCoy, K.F., Paliouras, G. (Eds.). User Modeling (UM 2007). Springer Lecture Notes in Computer Science 4511, 2007. [8] Denaux, R., Dimitrova, V., Aroyo, L. Integrating open user modeling and learning content management for the semantic web. In Proceedings of UM 2005, User Modeling (2005) 918. [9] De Bra, P., Houben, G.J., Wu, H. AHAM: A Dexter-Based Reference Model for Adaptive Hypermedia. In Proceedings of ACM Hypertext 1999 (1999), 147-156. [10] Ghosh, R., Dekhil, M. Mashups for semantic user profiles. In Proceedings of WWW 2008 (2008), 1229-1230. [11] Grapple project: www.grapple-project.org [12] Heckmann, D., Schwartz, T., Brandherm, B., Schmitz, M., von Wilamowitz-Moellendorff, M. GUMO - The General User Model Ontology. In Proceedings of UM 2005, User Modeling (2005) 428-432. [13] Houben, G.J., McCalla, G.I., Pianesi, F., Zancanaro, M. (Eds.). User Modeling, Adaptation, and Personalization (UMAP 2009). Springer Lecture Notes in Computer Science 5535, 2009. [14] Kuflik, T. Semantically-Enhanced User Models Mediation: Research Agenda. In Proceedings of UbiqUM'2008, workshop in conjunction with IUI 2008 (2008). [15] Linked Data: www.linkeddata.org [16] Nejdl, W., Kay, J., Pu, P., Herder, E. (Eds.). Adaptive Hypermedia and Adaptive Web-Based Systems (AH 2008). Springer Lecture Notes in Computer Science 5149, 2008.

6. REFERENCES
[1] Abel, F., Heckmann, D., Herder, E., Hidders, J., Houben, G.J., Krause, D., Leonardi, E., van der Sluijs, K. A Framework for Flexible User Profile Mashups. In Proceedings. of APWEB 2.0 2009, workshop in conjunction with UMAP 2009 (2009). [2] Aroyo, L., Dolog, P., Houben, G.J., Kravcik, M., Naeve, A., Nilsson, M., Wild, F. Interoperability in personalized adaptive learning. J. Educational Technology & Society, 9 (2) (2006) 418. [3] Bailey, C., Hall, W., Millar, D., Weal, M. Towards open adaptive hypermedia. In Proceedings of AH 2002, Adaptive Hypermedia and Adaptive Web-Based Systems (2002), 36-46. [4] Brusilovsky, P. Adaptive hypermedia. User Modeling and User Adapted Interaction, 11 (1/2) (2001) 87-110. [5] Brusilovsky, P., Kobsa, A., Nejdl, W. (Eds.). The Adaptive Web. Springer Lecture Notes in Computer Science 4321, 2007.

You might also like