ERCIM News No.46, July 2001 [contents]
by Judy Kay
We are seeing a huge growth in personalisation and customisation of interfaces. This introduces the need for a collection of information about the individual user. Within the adaptive user interface community, this is called a user model. This is any individual, dynamic and explicit model the system holds about the user. It might represent the users beliefs and preferences as well as other user attributes. The issue of scrutability of user models arises.
With the user model as the driving force for a systems adaptation to the individual user, we can illustrate the issues for scrutability of user models in the following scenario.
Jane starts Mynews, a personalised electronic newspaper. First she gets the headlines. Today, these include:
Mynews is supposed to select just those news items most likely to interest this user. It models Janes interests so that it can collect reports on issues she will want to know about. Where there are several items about an issue, the user model is supposed to ensure selection of reports Jane will prefer. For example, she may like a particular critics music reviews or one syndicates reports for European news. In this scenario, Jane might ask:
This scenario would probably need a large model of user interests and preferences. The first two questions relate directly to parts of the user model representing Janes preference for Flipper movies and Carr-Boyd music. The user should be able to find information to help answer the questions if they have access to the value of that part of the model and a record of the processes which contributed to that value. It may be that Jane has rated several Flipper movies and other animal-centred childrens movies, usually giving very positive ratings. She may have also ordered several such movies from an on-line service that provides such information to the user model. If such information affected the values in the user model, Jane should be able to scrutinise the model to see that this was so.
The third question might be answered within the articles of the newspaper. Generally, such questions require an explanation of part of the user model ontology, in this case, a music ontology with Carr-Boyd or Australian contemporary art-music composer. The fourth question relates to user control of the model.
The last three questions are likely to involve a combination of the user model and the mechanisms controlling Mynews. These questions relate to issues of media control.
Why might we want scrutability? The strongest reason is that users should be able to feel in control. When the user knows that a screen has been adapted to them, it is natural for them to wonder just what has been adapted to them. What might a different user see? Are they missing out on something? This view is supported by the 1995 European Community Directive on Data Protection (http://www.doc.gov/ecommerce/eudir.htm, visited May, 2000) which states that personal data for customisation requires an intelligible form of the data undergoing processing and of any available information as to their source.
We have been working in this area on several fronts. We have developed the accretion representation for user models. This ensures that the interface can explain what it believes about the user and why in terms of the sources of information and the way these are interpreted. From that foundation, we have been working on interfaces for viewing user models, both in detail as well as with visualisation interfaces which enable the user to gain an overview of a large user model. We also have projects which extend the scrutability beyond the user model and into the application. For example, in the context of interfaces to support learning we have explored a scrutable hypertext interface, and, in other projects, we are exploring a scrutable teaching strategy. In broader applications, we are building experimental systems that assist the user in managing documents with scrutable assistant programs.
Judy Kay University of Sydney, Australia
Tel: +61 2 9351 4502