Towards a General-Purpose Adaptive Hypermedia System

Paul De Bra^*+, Ad Aerts, Geert-Jan Houben⁺, Hongjing Wu
Department of Computing Science
Eindhoven University of Technology (TUE)
PO Box 513, Eindhoven
The Netherlands
{debra,wsinatma,houben,hongjing}@win.tue.nl

This paper resides at http://wwwis.win.tue.nl/~debra/edmedia2000/shortpaper.html.

Abstract: This short (work in progress) paper describes recent updates to the "Adaptive Hypermedia Architecture" (AHA), developed at the Eindhoven University of Technology (TUE). AHA has been used for two courses at the TUE, and experimentally also in a kiosk information system. The next set of applications of AHA consists of an adaptive geological atlas and an adaptive and personalized version of the (central) university's website.

Keywords: adaptive hypermedia, user modeling, adaptation rules

Introduction

Adaptive hypermedia applications [B96] exist in many different forms. However, this research and development area is largely dominated by systems that have been used for a single educational application. Notable exceptions are Interbook [BSW96] and AHA [DC98a,DC98b]. Both systems have been used to implement different applications (still in the area of education), and are also used in tutorials on adaptive hypermedia, including those at the WebNet'98 and the ED-MEDIA'2000 conferences.

Many educational Web-based systems share a common property: the system monitors the user's progress through an electronic textbook and gradually starts to offer more and more complex and detailed information. It maintains a user model that consists of knowledge about concepts. Each time a user visits an information page or completes a multiple-choice test the user model is updated. Typically each user action results in the "knowledge" about some concept to be augmented. In AHA the user model always consisted of Boolean variables. (The user knows or does not know a concept.) Interbook supports a few knowledge values, called not known, learned, well learned and well known. Yet other systems (including [PDS98]) use many more values, like a percentage (integer between 0 and 100).

In Section 2 we describe how AHA was updated to allow a richer set of knowledge values and to support non-educational applications as well. Section 3 briefly indicates how potential user model update problems, resulting from the added functionality, are avoided.

Making AHA more "General-Purpose"

The (old) basic functionality of AHA was the following:

• A page access or multiple-choice test results in a Boolean (concept) knowledge value becoming true.
• Fragments of content (text) can be conditionally included. The conditions are Boolean expressions using concepts.
• Links are shown in different colors, depending on Boolean conditions using concepts.

Recently AHA has been updated in three ways:

1. Instead of Boolean values the knowledge values in AHA are now integers between 0 and 100.
2. Each page access can augment or decrement the knowledge value of several "concepts". (So "knowledge" can decrease as well as increase.)
3. When a concept's knowledge value is altered a rule, associated with that concept, determines which other concepts' knowledge values are also (recursively) altered.

These changes in the user model and how it can be updated have two important consequences:

• The "knowledge" value of a concept can now truly be used to mean arbitrary other aspects of an application because it is just a numerical value that can be augmented or decremented. (The evolution of "knowledge" is no longer monotonic.) This opens more possibilities to use AHA in non-educational applications or to implement adaptive features that are not related to knowledge.
• It becomes possible to associate several pages with a single composite concepts. Accesses to pages can contribute towards the knowledge of a composite concept. It also becomes easy to express the requirement that reading a (large) subset of these pages is sufficient to consider the composite concept as "known".

A typical new possibility of the non-monotonic evolution of knowledge values is opening and closing menus. In AHA one can create a "table of contents" frame in which chapters are shown plus the section-structure of the "current" chapter:

• chapter 1
  • section 1.1             shown if chapter 1 is current
  • section 1.2
  • ...
• chapter 2
• ...

Such a menu structure is implemented by making the submenus content fragments that are shown under the condition that the value of the corresponding chapter "concept" is 100. When chapter X is selected, the other chapters are set to 0 and chapter X is set to 100. In the "old" AHA it was not possible to decrease knowledge values, and hence such opening and closing of menus could not be implemented.

Update Problems in AHA

Unfortunately, the versatile propagation of knowledge using update rules in AHA leads to some undesirable run-time behavior:

• The propagation of knowledge can cause one page access to induce several updates to one and the same concept. Access to A can imply an update to B and C and both B and C may induce an update to D. It is not a priori clear whether such an effect is desirable or acceptable.
• It is easy to (inadvertently) cause infinite recursive update loops. In the example of the (sub)menus, the update to the "knowledge" value of the menus that are being closed may inadvertently induce a new update to the menu that is being opened. When augmenting A by X means decrementing B by X it also means that decrementing A by X means incrementing B by X. When A updates B and B updates A, an infinite loop is imminent.

There are three easy ways to avoid infinite loops:

1. Disallow the definition of update rules with loops. (Loop detection in these definitions is easy.)
2. Allow each concept to be updated only once (per page access).
3. Allow each update rule to be used only once (per page access).

Each of these measures has its merits and drawbacks:

1. Circular definitions in update rules are sometimes desirable, like in the submenu example. Each menu excludes the other menus and the update rules to implement this are circular (by nature).
2. When we allow only one update to a concept (per page access) the example in which A updates B and C and B and C each update D becomes non-deterministic. (Updating D through A and B may result in a different change to D than updating it through A and C.)
3. If we allow each update rule to be used at most once (per page access) then the update algorithm becomes deterministic and infinite loops become impossible.

Unfortunately, one can easily see that in the case of the submenus the third (and only seemingly acceptable) loop-avoidance method does not result in the correct menu being opened. This is a consequence of propagating negative updates. Therefore AHA uses the following method:

4. Each update rule is used at most once, but in addition, when a rule for a concept contains a negative value it is not applied in recursive steps that decrease the value of that concept. (In other words, we do not allow recursive steps with a double negation.)

Conclusion

AHA has been made more general-purpose by making the user model more flexible, with a rich knowledge value set and non-monotonic behavior. Recursive user model updates are implemented in a safe way which results in deterministic and finite behavior. New applications of AHA make use of adaptive features that are not related to user knowledge.

References