Lisa Tweedie
Department of Electrical and Electronic Engineering,
Imperial College of Science, Technology and Medicine
South Kensington , London, SW7 2BT
Tel: +44 171 594 6261
l.tweedie@ic.ac.uk
Bertin [4] was probably the first to suggest matrix
construction as a way of examining IVAs. This involves
describing the data as a reorderable matrix of objects and
attributes. This is similar to the relational data model used
widely in computer science. Bertin argued that to be
effective IVAs must enable both attribute and object
focused questions to be asked. Additionally he identified
three distinct levels (FIGURE 1: Bertin's 3 levels) for these
questions: elementary (about a single item), intermediate
(about a group of items), and overall (about all the data).
FIGURE 1
FIGURE 1: Bertin's 3 levels
The object/attribute matrix forms the framework for our
DIVA notation (FIGURE 2: Object/Attribute Matrices). A
heavy line across the grid distinguishes different types of
data. Two forms of distinction are made: imposed
distinctions (e.g. houses I like and houses I don't) and
inherent distinctions (e.g. the deterministic relation
between parameters and functions found in many
modelling problems).
FIGURE 2
FIGURE 2: Object/Attribute Matrices
Consequences are the effects that a data manipulation has
on the data. These effects can be:
FIGURE 3
FIGURE 3: The Attribute Explorer (left) and the Dynamic House finder (right)
The simplified DIVA description in figure 4 (FIGURE 4: a
simplified DIVA description) shows that the data for both
IVAs has two inherently distinct attribute types: ordinal
(sliders) and nominal (buttons). In the Dynamic
HomeFinder the data shown is the intersection of all the
selections currently made (presentational operator). In
order to elicit the effect of changing an attribute the user
has to make a Dynamic Comparison. The Attribute
Explorer uses a more Venn-diagram-like representational
operator. This allows the set relations between two
attributes to be compared both statically and dynamically.
FIGURE 4
FIGURE 4: a simplified DIVA description
The numbers on the Attribute and Object axes indicate the
level of question that can be asked (1 = elementary, 2=
intermediate, 3= global). The two IVAs allow attribute-bin
selection at all three levels but, on the ordinal scales
(sliders) this is limited to one selection (at any level). Both
IVAs only allow object selection at the intermediate and
global level. In other words individual objects cannot be
selected. To overcome this house lines connecting a house's
squares were an additional tool in the Attribute Explorer.
Both IVAs have the following shortcomings: the single
selection limit on the sliders prevents disjunctive selection
of the ordinal data; individual objects can't be selected; the
limited use of presentational operators means that the user
has little flexibility in displaying the data. The difference
betwen the two IVAs lies in their consequence operators. It
seems that representational consequence operators facilitate
static comparison. These ideas were not evident to me as a
designer [7] prior to using DIVA.
Despite this drawback I have found that using DIVA does
inspire novel design ideas. Bertrand Russell is quoted as
saying "a good notation has a subtlety and a
suggestiveness which make it seem at times like a live
teacher". My hope is that DIVA can evolve into such a
teacher.
Abstract
DIVA is a notation for describing interactive visualization
artifacts (IVA). This notation forms one part of my thesis
work - the overall aim of this thesis is to find ways to
improve the design of IVAs. By describing different IVAs I
hope to elicit general principles to aid this process.
Keywords:
Visualization, Interactive Graphics
Introduction
The advent of powerful graphical computers means that
interactive visualization artifacts (IVAs) are a possibility.
These artifacts allow exploration and manipulation of data
with immediate visual feedback and have radically
different properties from traditional static graphs. The
questions asked here are: Can we describe these IVAs ?
Can we compare two artifact descriptions? What
conclusions can be elicited from such comparisons ? In this
paper I will outline a notation for such descriptions called
DIVA.
DATA MODELS AS A BASIS FOR DESCRIPTION
Benyon [1] has advocated using data models as a basis for
descriptions of human/artifact systems. As he puts it: "Data
is ...probably the only thing humans have in common with
computers". Green [2] has put this idea into practice in the
form of structure maps (ERMIA diagrams) which try to
capture the relationships between information entities.
Robertson [3] has also matched representations to tasks by
matching data types (nominal or ordinal) and the level of
focus (see Bertin's "three levels" below).
PERCEPTUAL COMPARISONS
Visualization tasks can be characterized as perceptual
comparisons of one set of objects with another - these
comparisons can be between any pair of Bertin's levels.
However if we were to describe an IVA in terms of all the
perceptual comparisons that can be conducted, we would
hit a problem. A user can look at any part of a screen and
compare it with any other part. Any attempted formalism of
this process would result in a very general description.
What actually interests us are the perceptual comparisons
that an IVA specifically facilitates. This occurs in two ways
- first via the data manipulation tools that act on the data
and, second, via their layout which groups items together
spatially. These two aspects of a visualization may facilitate
the same perceptual comparisons or different ones. Thus
DIVA has three sorts of operator :
a) Data Manipulation operators:
A data manipulation
operation has two parts, an action and a consequence:
Actions : are events that the user initiates - this is covered
by the operator "selection". Selection can be attribute or
object focused and can occur at any of Bertin's levels.
b) Structural operators:
So far only group and order
operators have been used. There are probably others.
c) Perceptual operators:
The basic perceptual operator is
comparison. There are two types - static comparison
(concurrent) and dynamic comparison (consecutive). An
assumption that is inherent in the DIVA notation is that one
can only compare two sets at once although this can occur
between any of Bertin's 3 levels.
EXAMPLES
Simplified versions of two similar IVAs will be used as an
illustration (the structural operators have been omitted due
to lack of space). In both cases the task is house search and
attributes of the house data are assigned to scales. Both
IVAs use sliders and buttons to select the data (FIGURE 3:
The Attribute Explorer (left) and the Dynamic House finder
(right)). In the "Dynamic HomeFinder" [6] these selections
are used to specify the population of houses that show up
on a map. In the "Attribute Explorer" [7] histograms are
positioned to the right of the scales. The histograms have a
square for each house in the data . When a range is
selected, those same house's icons are highlighted on each
of the other scales. If a range on a second scale is selected
(with a second colour) these houses are also highlighted. If
a house satisfies both requirements the two colours blend.
In figure 3 this blending is indicated using crosses.
SUMMARY
DIVA does have a major inadequacy: when we visualize
we use both perceptually initiated (opportunistic)
processing and cognitively initiated processing (planning),
as it stands DIVA only focuses on the former. A
description of an IVA's semantics might give a designer
clues about the cognitive effort required to understand it. In
this context ERMIA [2] diagrams could usefully
complement DIVA.
ACKNOWLEDGEMENTS
This work is funded by a grant from EPSRC (UK) and
sponsorship from Philips Research, Redhill. Thanks to Bob
Spence and Thomas Green for their advice and support.
