The development of novel interactions based on L-systems [30] represents an important step
toward the application of the iconographic technique for exploring complex and multidimensional data sets.
There was a need in the iconographic technique for useful ways to interacting with the visualization. One key
point in all of these interactions is that they can provide a qualitative idea about the distribution of values in
the data. This fact is particularly important in the iconographic technique since one of the problems associated
with the iconographic technique is precisely the lack of ways of figuring out what is the relationship between
patterns in the iconographic display and structure or relations in the data.
This course of research will develop new interaction techniques oriented either to provide a more
quantitative measure of the relationship behind revealed patterns or to provide support for further visual
exploration. New interaction techniques are needed to support learning about the underlying data and its
relation to the iconographic representation of this data.
There are several things that could be useful to users exploring databases with the iconographic technique. For
example users could be interested in knowing how a different mapping could change the visual
representation of some particular region of interest. A simplified approach to provide this kind of facility
may be the development of an interaction that changes the mapping in the selected region based on a
predefined criterion. For example, a criterion could be swapping the attributes mapped to limbs that are on
opposite sides. More complex criteria could be specified based on, for example, data statistics in the
selected region. The mapping in the selected region could be determined by the correlation between data
attributes. For example attributes showing higher correlation could be mapped to icon parameters affecting
the same geometric attribute (length, angle, or color). This thesis will explore the implementation of this
kind of interaction and the identification of high level criteria capable of providing useful information about
the effect that a particular mapping has over the iconographic representation.
Figure 5: A interaction that dynamically changes the mapping in the iconographic display
Figure 5 shows an approach to implement the mapping interaction. The approach is based on the 2D
moveable box described in the previous section, but in this case a function for changing the mapping in the
region defined by the moveable box has been selected. A criterion for changing the mapping must be also
specified. Moving the moveable box over the iconographic representation will change the mapping in the
area defined by the moveable box according to the specified criterion. The small area in the low right corner
of the window shows the iconographic representation using the new mapping. A textual indication of the
original mapping and the new mapping is also shown. Since the criterion specified for changing the mapping
not necessarily determine the same mapping for different regions of the iconographic display, this
approach can provides the basis to investigate how a particular data structure resonates with a particular
mapping.
Another thing that users could be interested is the effect that normalization has over the visual
representation. Systems using the iconographic technique (L-system, Exvis) normalize all the attributes values
using local normalization. In local normalization each attribute is normalized in such a way that its
minimum value correspond to zero and its maximum value to one. Local normalization is very useful in
dealing with attributes whose values span along ranges that can differ in one or more order of magnitude,
because it allows that each attribute span along the whole normalized range. The problem with this
normalization approach is that normalized attributes can not be directly compared because they are
affected by a different scale factor.
On the other hand global normalization uses the absolute minimum and maximum values of the whole data
set to normalize each attribute. Thus, each value in the data domain is represented by a unique value in the
normalized domain, enabling the direct comparison of attributes in the normalized domain. The visual
representation is also affected by this normalization. Some patterns can be made apparent by the
normalization approach used. For example, a pattern can be created by the fact that all the minimum values
are mapped to the same normalized value in a local normalization approach. An interaction could be
developed to allow users to interactively change the normalization approach in selected regions of interest
in order to determine if a revealed pattern is a consequence of the normalization approach used or if it
actually represents a structure in the data. A local normalization can provide a qualitative idea about the
data driving the icon parameters, but a global normalization could provide a more quantitative idea about
the actual values driving the icons parameters.
Users could also be interested in determining, in a qualitative or quantitative way, how attributes are
changing along the dimensions being used for driving the axes in an iconographic display. In some cases trends
along specific directions or localized areas can be easily discerned by the visual system but in other cases
these trends can be difficult to discern. Several factors could be related to this difficulty. The mapping
between attribute values and icon parameters, the attributes used for driving the axis, and global
parameters affecting the icon geometry could be related to this difficulty. The iconographic technique could
benefit from a mechanism able to provide an enhanced view of the iconographic representation in such a way
that trends along different directions can be visually discriminated. Figure 6 shows an iconographic representation of the FBI homicide
database from 1985 to 1987. In this visualization "AGEV1" (age of the victim) is driving the X axis
and "AGEF1" (age of the offender) is driving the Y axis. "NVICT" (number of victims) is driving the limb
2 (angle, length, and intensity) and "NOFF" (number of offenders) is driving the
limb 4. The limb 1 is
controlled by "SEXF1" (sex of the offender), "YEAR" (year of occurrence) and "MON" (month of
occurrence). The limb 3 is controlled by "SEXF1", "YEAR", and "NOFF".
Figure 6: An iconographic representation of the FBI homicide database
In image processing there are operations that by operating on a group of pixels are able to raise or reduce specific attributes of one
image. For example, high spatial frequencies (gray level changing rapidly along any direction) can be reduced by using a low pass filter.
There are also spatial filters that allow edge enhancement in an image [6]. For example the
Prewitt gradient operation is able to enhance borders along specific directions and the Laplacian edge
enhancement is able to highlight edges in an image, regardless of their orientation. This research will
explore the extension of these images operations to the iconographic technique in order to provide other views of
the iconographic representation. In these views specific trends in the data would be emphasized and hence
would be more easily discerned by users. Figure 7 shows the same iconographic representation shown in
Figure 5, but in this case a spatial low pass filter was applied.
Figure 7: A low pass filtered iconographic representation of the FBI homicide database
This thesis intends to use this approach to identify possible regions of interest which, once
identified, could be used as starting points for further exploration. For example, the identified region could
be selected and quantitative values from that region could be determined and visualized using another
visualization technique.