Proceedings of the 2nd Workshop on Adaptive Hypertext and Hypermedia,
HYPERTEXT'98, Pittsburgh, USA, June 20-24, 1998

Modelling Personalizable Hyperlink-Based Interaction

James Ohene-Djan, Alvaro A.A. Fernandes
Department of Mathematical and Computing Sciences
Goldsmiths College, University of London
New Cross, London SE14 6NW, UK
{ j.ohene-djan | fernandes }
Abstract : This paper describes the development of an abstract model for hyperlink-based interaction in which personalization can be studied with greater conceptual clarity than is possible by technology-driven experimentation. The model characterizes a rich set of abstract user-initiated tailoring actions, which enable individual users to come closer to satisfying their specific, and often dynamic, information retrieval goals. The model forms a foundation for our current work, a systematic investigation of the nature, scope and effects of system-initiated tailoring actions on hyperlink-based systems(HLBSs).

Keywords: Hypermedia Design, WWW Personalization, Hypermedia Modelling.


Research into Personalization & adaptation(P&A) actions in HLBSs is motivated by a great interest (both scientific and commercial) in increasing the effectiveness of HLBSs as a platform for information retrieval tasks in which different users have different information goals and different histories. In HLBSs that lack P&A actions most of the interaction a user might experience with a hyperdocument is determined by the design decisions that shaped the hyperdocument in terms of its content, its rendering aspects and the navigation possibilities it offers the user. As a consequence the designer owns the hyperdocument. Thus, in HLBSs that lack P&A actions, users can navigate through a hyperdocument using links, but they are prevented from enforcing their individual preferences as to content, rendering and navigation possibilities.

Our approach to overcoming this impediment is to extend HLBSs with P&A actions that effect a transfer of ownership from the original designers of the hyperdocument to each of its users, thereby enabling the latter to redesign the former according to their specific information goals and histories.

We also draw some of our motivation from the fact that there seems to be no discernible consensus among researchers in adaptive hypermedia with respect to following questions:

  1. Which are the emergent properties of HLBSs? Equivalently, what is the scope of P&A in HLBSs?
  2. Which P&A actions could be made available to users? Equivalently, what descriptive stance should be taken with respect to P&A actions in HLBSs?
  3. Which P&A actions should be made available to users? Equivalently, what prescriptive stance should be taken with respect to P&A actions in HLBSs?
We would argue for a precise, abstract characterization of what emergent properties can be assigned to HLBSs, so that issues relating to P&A in the technologies of which an open HLBSs tends to be a client (e.g., in database and in user-interface technologies) are not confounded with P&A issues in HLBSs. In other words, we believe it is important to characterize, at an informative level of abstraction, what is unique to HLBSs rather than inherited (or shared) with server technologies. Such a characterization would open the way to a principled exploration of specific P&A issues in HLBSs.

Our Approach

The main ideas guiding our approach can be phrased as follows:
  1. The model is an abstract model, as many steps removed from concrete implementations as necessary to allow as systematic, exhaustive investigation of P&A issues in HLBSs.
  2. The model is an open model, insofar as we view HLBSs as clients of a variety of servers, an in particular of data and user-interface servers.
  3. Personalization involves a transfer of ownership of the process of interaction with a hyperdocument, from designers to users.
  4. To ensure that the set of personalization actions is consistent, its elements are induced from the formal definition of the hyperdocuments they act upon.
  5. All design decisions are, in principle, in scope for personalization actions.
  6. Our model of hyperlink-based personalization can express most, if not all, personalization actions proposed in the literature (see[1] for a comprehensive review).
  7. Our model describes which personalization actions can be made available. In order to prescribe which should be made available, empirical studies are needed.

THE Goldsmiths Hyperlink Model

To model adaptive, personalizable hyperlink-based interaction we propose a model core hyperlink behaviour by partitioning it into three regions. Non-adaptive, non-personalizable HLBSs are modelled by the functions provided by what we refer to as the H-region. Personalizable HLBSs require the addition to the H-region of the functions provided by what we refer to as the P-region. This causes no disruption whatsoever and requires no changes at all to the H-region. Adaptive HLBSs require the addition to the P- and H-regions of the functions provided by what we refer to as the A-region.

The H-Region

The H-region is formalized as a composer from specifications, i.e., what a designer writes is not a document but rather a specification of how to build the document upon request. A formal language for writing such specifications has been defined along with a formal abstract machine to execute them thereby yielding renderable documents. At the most basic level, the functionality provided by the H-region is to process specifications of hyperdocuments into renderable texts. In this approach, a designer writes specifications as to where the contents can be found, what rendering is the document to have and how to compose content and presentation features into a renderable text that a user sees. The basic dynamics of the H-region is the following. The user requests a page to be rendered, as usual. Such a page exists, and is fetched, as a specification of where to find its content and how to render it. One then proceeds to fetch the contents by client-server querying designer-specified sources. The retrieved content is then composed, as specified by the designer, into the text to be rendered. Finally, the core responds to the request with the text thus composed.

The P-Region

The P-region comprises a group of functions that are non-disruptively added to the H-region in order to model personalizable hyperlink-based interaction. The P-region provides two basic processes: the personalization of hyperpages by annotation and rewriting, and the enforcement over a renderable text of previously expressed preferences (in the form of notes on a hyperpage). When superimposing the P-region onto the H-region, users can not only request a hyperpage, but also annotate or rewrite it, thereby creating their own version of it. Thus design of that hyperpage can therefore be overridden by user and this event characterizes ownership transfer.

The kinds of personalization actions that we model are based on annotating and rewriting the hyperpage specifications. Annotation pairs a hyperpage specification with notes of interest to the user, and by doing so, presumes that versioning takes place. Such notes take one of the following forms. Firstly, a note can assign user-specific values to user-generic attributes of interest (e.g., that the level of difficulty of a given page is high, or that `planets' is a keyword of relevance to a given segment of a page). Secondly, a note can specify a rewriting action over the renderable text after it has been composed by the H-region, i.e., after content has been fetched and made ready for display (e.g., to map American into British spelling forms). This form of post-composition rewriting can also be conditional on the environment (e.g., replace images with captions if the display unit is text-only). A formal language for tailoring the hyperdocument specified by the designer into a personalized version has been defined along with a formal abstract machine to generate them. For details, see[3]. The existence of annotations on hyperpages allows for:

  1. personalization of a specified hyperpage;
  2. the specification of alternatives to a specified hyperpage;
  3. the specification of comparable hyperpages to a specified one; and
  4. the recording of information about a hyperpage (i.e., what are the current values of attributes set by previous annotations).

The A-Region

The A-region comprises a group of functions that are non-disruptively added to the P-region to model Adaptivity. We view adaptation as system-initiated personalization. The immediate import of this is that our conception of adaptation requires a HLBS, at least, to employ a model of each user as the basis of a prescriptive theory of what personalization action might the user have taken, and to initiate that action on behalf of the user. Furthermore in our view, adaptation is, in principle, as expressive as personalization and requires no other technologies than those involved in user modelling and in decision making from a user model. The A-region enables users and designers to define strategies as to when the system should take the initiative and actively tailor the interaction to a user in the light of that user's information goals and history of use.

Conclusions and current work

The approach we are currently exploring to add, adaptive capabilities centers on an adaptation function. This function implements an inference engine over a decision theory (i.e., a theory as to which actions are more likely to yield the most benefits given some accumulated knowledge of past interactions). The accumulated knowledge are the information goals and the history of each user, while the actions which the inference engine is in charge of suggesting are personalization actions as defined by the P-Region.[2] Besides greater conceptual clarity regarding the scope for P&A in HLBSs. The model summarized in this paper we hope represents a contribution towards a better foundation to underlie the exploration of P&A[1] issues in HLBSs.


[1] Peter Brusilovsky. Methods and techniques of adaptive hypermedia. User Modeling and User-Adapted Interaction, 6:87--129, 1996.

[2] A.A.A. Fernandes, M.H. Williams, and N.W. Paton. A Logic-Based Integration of Active and Deductive Databases. New Generation Computing, 15(2):205--244, 1997.

[3] James Ohene-Djan and Alvaro A.A. Fernandes. Personalizable hyperlink-based interaction. Technical report, Department of Mathematical and Computing Sciences, Goldsmiths College, University of London, 1998.