How Can Users Edit and Control their Models in Ubicomp Environments

Abstract

We describe a visualisation for user models that could be applied to ubiquitous computing environments. The visualisation allows users to explore their user model and scrutinise the components in models with hundreds, perhaps even thousands of components.

Full text

How can users edit and control their models in Ubicomp
environments?
Judy Kay, Andrew Lum, James Uther
School of Information Technologies
University of Sydney, Australia 2006
{judy,alum,jimu}@it.usyd.edu.au
Abstract
We describe a visualisation for user models that
could be applied to ubiquitous computing
environments. The visualisation allows users to
explore their user model and scrutinise the
components in models with hundreds, perhaps
even thousands of components.
Keywords scrutability, ubiquitous computing,
user modelling, visualisation.
1. Introduction
Ubiquitous computing environments are
characterised by the range of sensors, of varying
reliability, with information about the user.
Typically, sensors have limited computational
power and limited and potentially unreliable
communication with centralised services. They
may also operate in quite diverse ways. This
means that information for user modelling in
particularly prone to noise, uncertainty and
unreliability.
At the same time, there may be similar problems
of power, reliability and bandwidth into the
devices which can interact with the user in the
ubiquitous environment. On the other hand, there
may be possibilities for rich interaction using
gesture, projected images that the user can
interact with, and interactive wall displays of
other types. It is more common to consider the
need for interaction with restricted devices such
as PDAs and phones. When, and if, the user
wants to understand the personalisation in such
an environment, they will need to be able to
scrutinise their user models with the aid of the
devices available.
If a user model is very simple and small, there
may be many possible ways to provide access to
the model and allow the user to alter it. It may be
that quite small parts of the user model might be
identified as relevant for the action of the
personalised part of the ubiquitous environment.
However, if the model is large, it becomes quite
challenging to display relevant parts of the model
effectively. One appealing approach is to use a
visualisation tool so that the user can readily see
and interact with a large number of user model
elements.
In the next section, we describe one approach to
this problem, a visualisation designed to display
user models. In the subsequent section, we
discuss its applicability for ubiquitous computing
environments, both within them and in
conjunction with them followed by a discussion
and conclusion.
2. VlUM visualisation
Visualisation of Large User Models (VlUM) is a
program written to display large user models in
web-based systems (Uther, 2001). It was inspired
primarily by Murtagh’s work on automated
storytelling systems (Murtagh, 1996). VlUM is
implemented as a Java Applet, which occupies a
frame on one side of the browser window. It
utilises perspective distortion to allow users to
navigate the user model. Users can see the whole
user model at once. Colour is used to show the
values of the components of the user model. The
user can see the structure of the model because it
is reflected in the font size, spacing and
brightness of the components.
VlUM also facilitates the viewing of multiple
datasets. The visualisation enables a user to see
their user model in absolute terms or in terms of
another standard. It also supports visualisation of
the user model in comparative terms. For
example, in the learning domain of its initial
development, students could see how their user

Figure 1: VlUM displaying a user model from a movie recommendation service. There are over 300 terms on the
display. Lawrence of Arabia is currently selected. The system is fairly certain the user will enjoy this movie.
model compared with that of the teacher’s
expectations or the whole class.
VlUM user model components have a score and
certainty. The score represents a heuristic value
for the domain. The certainty value shows the
system’s confidence that the score is accurate.
VlUM uses colour to provide an indication of a
component’s score through a slider that adjusts
the viewing standard. Any components with a
score less than the standard will have a red hue.
Ones greater than the standard will have a green
hue. The further a score is away from the
standard, the more saturated the colour will
become.
Clicking on a component with the mouse will
select it and put it into focus, and the display will
change so that the selected component will then
have the largest font. Other components will
have increased sizes the more relevant they are to
the selected one. Components that are not
relevant are dimmed and shown in a small font.
In Figure 1, we show an example of a VlUM
display for a user’s movie preferences. Note that
the actual display is larger than the one above
and the use of colour and animation is quite
important for the effectiveness of the user
interface that VlUM presents. In this figure, we
can see that Lawrence of Arabia is currently
selected, which is why it is larger than other
movie titles and has more space around it. It is
the most visible component of the user model. Its
most closely related movies are the next most
visible, being smaller and having less space than
Lawrence of Arabia, but larger and with more
space than the less closely related movies. These
include The World in His Arms which is about a
quarter of the way from the top of the display,
Land of the Pharaohs which is a little below it
and The Salamander, which is quite near the
bottom of the display.
VlUM addresses Shneiderman’s classic
visualisation tasks (Shneiderman, 1996) in the
following way:

• Overview: the entire user model is shown
on the display.
• Zoom: users can focus on a particular
element in the graph.
• Filter: non-relevant items are hidden with a
smaller font and darker colour.
• Details-on-demand: can find out more
about the focused item from the menu.
• Relate: related items are shown in a
progressively larger font the more relevant
they are.
• History: users can step back through
previously focused items.
• Extract: a search function is provided, that
puts in focus items matching the search.
VlUM is distinctive in that it was designed
explicitly for the purpose of displaying user
models and supporting visualisation of a large
number of components of a user model. VlUM
was extensively evaluated to assess how large a
user model could be navigated effectively by
users. It proved to perform well with up to the
largest data set in the evaluation, consisting of
700 concepts.
Any user modelling representation that can be
mapped to a graph has a natural representation in
VLUM. This means that it should be able to
handle such diverse representations as Bayesian
models or simple overlay models. The usability
evaluations on VlUM were essentially overlay
models of user preferences for movies.
3. Related Work
The VlUM work appears to be the only work to
date on providing an overview of substantial
sized user models. Moreover, it is among the
quite small set of visualizations which has been
carefully evaluated to assess whether users can
perform the goal tasks and to determine how this
performance is affected by the number of
elements displayed.
There are tools for visualising user models that
could possibly be adapted for use in ubiquitous
computing. For example, qv (Kay, 1999) also
gives an overview interface for um user models.
The interface displays the user model as a tree
hierarchy, with the root node on the left of the
screen. Subsequent nodes can be expanded or
collapsed. The user model represented user
knowledge about editors such as EMACS, SAM,
and vi. Branches that are not relevant to the
user’s current level of knowledge are initially
collapsed to reduce cognitive overload when
they use the interface to scrutinise their user
model. Different shapes indicated beliefs
(diamonds), knowledge (squares), and non-leaf
nodes (circle). The fill colour of the shapes
represents component values. Although qv was
designed to give an overview of a user model, it
uses a very different approach and was explicitly
designed to make quite modest numbers of
elements visible at any one time. Users can only
access the finest details of information by
navigating down the hierarchy.
From the perspective of providing a user model
overview, there are some very important
differences between qv and VlUM. Consider the
case where the use is currently focusing on a
concept such the modeled preference for the
movie qv. Figure 1 shows the VlUM display.
The corresponding display for qv would show an
hierarchical tree expanded to show this
component. This works well where there is a
natural hierarchical structure as in the case of
knowledge of text editors, the main domain of
the qv work. However, it is less suited to a
largely non-hierarchical domain such as movie
preferences. Even more significant is the
differences in the way these overview tools
enable the user to see outliers in the user model.
In the case of qv, the collapsed parts of the
model are displayed with a value that
summarises the average of the values under that
node. If there are many nodes with just one
having a different value, the user would be
unable to distinguish this from the case where all
the subsumed nodes were of the same value. By
contrast, in VlUM, suppose that almost all of a
model is one value, say true. Then almost all the
components are displayed in green. Even a single
false component will stand out in red. In
practice, this may be extremely important,
especially in the types of context that apply in
ubiquitous computing environments.
VISNET (Zapata-Rivera and Greer, 2000,
Zapata-Rivera, Neufeld et al., 1999) is a
visualisation designed to help people understand
Bayesian Belief Networks, which can be
naturally represented as network structures.
Arrows between nodes represent cause and the
nodes are effects. A tree is displayed with
distance between nodes indicating the strength of
the causal relation, and the nodes size and sizes
colour representing the belief values.

4. Applying VlUM in ubiquitous computing
environments
Within the current experimental ubiquitous
environments, with very restricted I/O devices,
the current VlUM would be difficult to use. This
still leaves the possibility that the user could
interact with the visualisation on a conventional
terminal. The possibility of doing this may be
sufficient to make the user confident about
allowing the model to be used for personalisation
within ubiquitous environments.
Figure 2: VlUM modified for small displays
showing the same model as Figure 1. A depth cut-
off value of 2 has been used to reduce the amount of
visible text in the visualisation.
Even within the constraints of PDAs and other
small devices, it may be possible to provide a
visualisation in the near term. We have
experimented with rendering the visualisation in
smaller resolutions. Figure 2 shows the same
applet rendered at 320x240 pixels, which is the
similar to many PDA resolutions. A limit to the
amount of expansion from a component has been
imposed on the spanning algorithm to reduce the
number of visible words on the display. The size
of the current applet is less than 400k.
5. Discussion and Conclusions
There is a need for users to be able to control
their user models. The European Union Data
Protection Directive (European Union, 1995)
states that a user’s personal data should be
accessible to them and be correctable if there is
incorrect data.
In a ubiquitous computing environment it is
especially important that the personal data that
constitutes user models should be scrutable (Kay,
1999) since much of the information will be
collected invisibly, even surreptitiously. These
user models are likely to be very large, perhaps
consisting of several sub-models representing
different contexts or services. The VlUM tool
allows users to see an overview of their complete
user model. This is a starting point to navigating
through the model, and scrutinising the finer
details.
This is especially important when users move
into a new environment and context. At that
point, the user may become aware of the
personalisation if it changes with the new
context. This may motivate them to see if the
information in their user model is relevant or,
perhaps, whether it is correct. The user may also
want to check whether aspects of the user model
should be available to that part of the
environment, before it is released and shared
with the environment. Users may also want to
propagate information to each other in everyday
tasks. For example, users may wish to share task
lists with each other to help coordinate work and
become more efficient as a group (Schneider,
Kortuem et al., 2000).
Previous work on user model visualisation tools
has involved conventional interfaces with large
screen, keyboard and mouse. We would expect
that these will need to be adapted to ubiquitous
computing environments. This follows work
such as the exploration of adaptation of core
interaction objects, such as menus, to smaller
displays such as PDAs (Rose, Stegmaiser et al.,
2003).
We have described a novel interface that
supports visualisation of a user model. This
serves as a foundation for users to gain an
overview of their user model and to navigate
through this so that they can edit and control that
user model. Effective user model visualisations
give users the understanding to control their user
model. Since this model is the core of control of
the personalisation, such interfaces are critical to
empowering users to control the personalisation
of their ubiquitous computing experience.

6. References
European Union, EU Directive 95/46/EC:
Protection of Individuals with Regard to
the Processing of Personal Data and on
the Free Movement of such Data.
Official Journal of the European
Communities. 1995.
Kay, J., A scrutable user modelling shell for
user-adapted interaction, 1999,
University of Sydney.
Murtagh, M., The Automatist Storytelling
System: Putting the Editor's Knowledge
in Software, Masters, 1996,
Massachusetts Institute of Technology.
Rose, D., et al. Non-invasive Adaptation of
Black-box User Interfaces. In: R. Biddle
and B. Thomas, Editors. Fourth
Australian User Interface Conference;
2003.
Schneider, J., et al. Disseminating Trust
Information in Wearable Communities.
In: 2nd International Symposium on
Handheld and Ubiquitious Computing;
2000.
Shneiderman, B. The eyes have it: A task by data
type taxonomy for information
visualizations. In: 1996 IEEE
Conference on Visual Languages; 1996,
p. 336-343.
Uther, J., On the Visualisation of Large User
Model in Web Based Systems, PhD
Thesis, 2001, University of Sydney.
Zapata-Rivera, J.D. and Greer, J., Inspecting and
Visualizing Distributed Bayesian
Student Models, in G. Gauthier, C.
Frasson, and K. VanLehn, Editors,
Intelligent Tutoring Systems. 2000. p.
544-553.
Zapata-Rivera, J.D., Neufeld, E., and Greer, J.
Visualization of Bayesian Belief
Networks. In: IEEE Visualization 1999;
1999.
Acknowledgements
We would like to thank Hewlett-Packard for
funding parts of this research, and Mr. Michael
Avery for proof-reading.