An Interactive 3D Integration of Parallel Coordinates and Star Glyphs

Elena Fanea∗ Sheelagh Carpendale∗ Tobias Isenberg∗

Department of Computer Science
University of Calgary, Canada

A BSTRACT with Star Glyphs. We show that, in fact, Parallel Coordinates and
Star Glyphs are orthogonal to each other and that they belong to a
Parallel Coordinates are a powerful method for visualizing multi- continuum of visualization techniques. We enrich this technique by
dimensional data but, when applied to large data sets, they become applying colour scales to the 3D Parallel Glyphs to support compar-
cluttered and difficult to read. Star Glyphs, on the other hand, can ison and selection tasks in 3D. In addition, we demonstrate that the
be used to display either the attributes of a data item or the values new interaction methods that are possible in 3D such as a ring ruler
across all items for a single attribute. Star Glyphs may readily pro- enhances comparison of data values and can lead users to entirely
vide a quick impression; however, since the full data set will require new insights into the data.
multiple glyphs, overall readings are more difficult. We present Par- The remainder of this paper is organized as follows. In the next
allel Glyphs, an interactive integration of the visual representations section we review previous work that is related to our own tech-
of Parallel Coordinates and Star Glyphs that utilizes the advantages nique. In the following section we introduce our concept of Parallel
of both representations to offset the disadvantages they have sepa- Glyphs and discuss the use of colour scales with the new visual-
rately. We discuss the role of uniform and stepped colour scales in ization. Section 4 addresses new interaction techniques that are
the visual comparison of non-adjacent items and Star Glyphs. Par- possible with Parallel Glyphs. Then, we present a short case study
allel Glyphs provide capabilities for focus-in-context exploration that illustrates the application of our technique to a specific data
using two types of lenses and interactions specific to the 3D space. set. Finally, we give a conclusion and suggest directions for future
work.
CR Categories: I.3.6 [Computer Graphics]: Methodology and
Techniques—Interaction Techniques; H.5.2 [Information Interfaces
and Presentation]: User Interfaces—Graphical User Interfaces, In- 2 R ELATED W ORK
teraction Styles.
Parallel Coordinates [6, 7] and Star Glyphs [14, 15, 18] are both
Keywords: Parallel Glyphs, parallel coordinates, star glyphs, well recognized multi-dimensional visualizations. In fact, there con-
multi-dimensional data sets, 3D visualization. tinues to be considerable research extending them and improving
the set of interactions possible with them [3, 12, 13, 18, 19, 25].
1 I NTRODUCTION Advancements for interacting with Parallel Coordinates include
brushing and multi-dimensional reordering to better support the in-
Parallel Coordinates are a powerful visual representation of multi- teractive exploration of multi-dimensional data [2, 12, 13, 18, 19].
dimensional data due to their capacity to display the entire dataset Brushing can be used to interactively select a range of values from
while supporting interactions that lead to identification of rela- a given dimension. This selected range is useful for filtering [25]
tionships among items and among dimensions [6, 7]. Several re- and discovering answers for particular queries [1]. In addition, fo-
searchers have been working on methods for improving aspects of cus + context views can be provided by expanding or contracting the
their capability to support visual analysis of multi-dimensional data. spaces between dimensions [25]. This can be important when the
The main motivation for developing a new visual representation is number of dimensions in the data set leads to a particularly dense
to address the problem of overlapping polylines. When two or more display [27].
items in the data have the same values for a subset of the attributes, Trends in data can be revealed in Parallel Coordinates using an-
the polylines drawn to show each item may overlap, depending on gular brushing [5]. With this interaction technique the user specifies
the order in which the attributes are plotted. This artifact frequently a subset of slopes to mark the corresponding data points as part of
occurs in Parallel Coordinates when even a very small number of the current focus. The subset resulting from this brushing can be
data items are shown. In this case it is hard to read the visualization used for filtering the initial set. Thus, the user can choose not to
to determine the values of dimension for a specific data item. In ad- display the items or dimensions that are relatively unimportant for
dition, due to the nature of Parallel Coordinates the order in which a particular task [25].
data items are plotted is lost. In cases where this order is meaning- A common interactive technique for reducing clutter and reveal-
ful we lose the ability to compare the development of data values ing data relationships in Parallel Coordinates is the manual or au-
along one dimension or to compare it with another dimension. In tomatic reordering of the dimensions [13, 18]. While manual re-
contrast, evaluating trends in the order of data values for one di- ordering may be more tedious than automatic reordering, it is often
mension is easy with the Star Glyph visualization. However, be- preferred because it allows the user to establish a superior reorder-
cause the Star Glyph visualization is spatially separate, direct com- ing [25]. Users can take specific aspects of the data into considera-
parisons between two glyphs are error prone. tion that might be overlooked by an algorithm that reorders based
In this paper we introduce Parallel Glyphs, a new visualization on similarity measures. Dimension reordering is also applicable to
to overcome these problems. Our technique extends 2D Parallel Star Glyphs as each spike in a glyph represents a dimension and
Coordinates into the third dimension and naturally connects them adjacent spikes can be easier to compare [25].
Another problem with the Parallel Coordinates technique arises
∗ e-mail: {faneael | sheelagh | isenberg}@cpsc.ucalgary.ca when it is used to visualize highly-dimensional data. This typically
results in visual clutter since plotted lines overlap extensively, mak-
IEEE Symposium on Information Visualization 2005 ing the visualization difficult to read. One approach to the miti-
October 23-25, Minneapolis, MN, USA gation of this problem is to use hierarchical clustering, which re-
0-7803-9464-X/05/$20.00 ©2005 IEEE. peatedly groups dimensions based on similarity measures. Thus,

149
 this representation provides a multi-resolution view of the data that item or a data dimension. Spikes, frequently bounded by a polygon,
reveals trends at different degrees of summarization. The same emanate radially from a central point. When a glyph represents a
idea of cluster-based hierarchy can be applied for both dimensions data item, its spikes vary in length according to the values the item
[9, 25, 26] and data items [3, 12]. Another technique, polyline av- has in each dimension. When, on the other hand, a glyph represents
eraging, which represents a range of polylines with an averaged po- one dimension, the spikes vary in length according to the value each
lyline, can be used for dynamically summarizing a set of polylines item has for that particular dimension.
[16]. It reduces the clutter arising from overlapping polylines and
is an alternative to more computationally demanding methods such
hierarchical clustering.
For hierarchically clustered representations X MDV T OOL [12,
18] implements n-dimensional brushing for interaction at a chosen
level in the hierarchy. Angular brushing can also be done automati-
cally or in relation to the hierarchy, if one is established [5].
Focus + context methods are general techniques developed to ad- Figure 1: Examples for Star Glyphs.
dress problems of visibility of details and have been applied to Par-
allel Coordinates [25]. Multi-focus distortion operations allow the
user to enlarge or shrink several regions in the same presentation
[26]. The space between dimensions can be increased or decreased 3.1 Adding the Third Dimension
with a simple click. If, with high-dimensional data, the space be- One significant motivation for developing Parallel Glyphs is to ad-
tween dimensions is too tight to allow an accurate click, a structure- dress the problem of overlapping polylines. This artifact frequently
based distortion called I NTER R ING can be used [27]. In an auto- occurs in Parallel Coordinates even when only a small number of
matic application of distortion, the spacing parameter is the distance data items are shown. For example, Figure 2(a) shows a dataset that
(for Parallel Coordinates) or angle (for Star Glyphs) [25]. Although contains only few data items. Looking at the right side of the Par-
the spacing between dimensions in Parallel Coordinates and Star allel Coordinate display in Figure 2(a) one can see four polylines,
Glyphs is typically uniform, non-uniform variations therein can con- indicating that there are four data items. However, on the left side
vey additional information about the dimensions [25]. only three polylines are visible. The 2D visualization of Parallel Co-
The various types of interactions possible with Parallel Coordi- ordinates can be ambiguous. For example, Polyline A is obscured
nates, Star Glyphs, Scatter Plots, and Dimensional Stacking have by either Polyline B or C between dimensions 1 and 3. This fact
been compared by several authors with respect to each of these four can affect even simple tasks such as revealing trends. By opening
visualization techniques [3, 12, 13, 18, 19, 25]. the Parallel Coordinates into 3D (see Figure 2(b)) we are able to tell
Several systems use Parallel Coordinates along with other visual not only that Polyline A follows the path of Polyline B but also that
representations. For example, Wong et al. [23] developed a 2D visu- there is in fact another polyline (D) following a different path that
alization system that contains Parallel Coordinates and Scatter Plots, was entirely hidden by the other polylines.
each presented in their own regions of the window. The represen-
tations are linked so that interactive changes to one are reflected in
the other. The two techniques combined better support data explo-
ration.
Tominski et al. [17] suggest organizing Star Glyphs on time
based axis. This would provide a linked 3D layout for Star Glyphs
with one glyph per data item but was left as a suggestion for fu-
ture work. Kiviat diagrams [4] are 2D visualizations that resemble
Star Glyphs in that radial spikes emerge from a common centre and
each spike corresponds to a data item. Kiviat tubes are obtained
(a) Regular Parallel Coordinates.
by displaying Kiviat diagrams along a time axis and rendering the
surrounding surface as suggested by Tominski et al. [17]. The sur-
face that connects all the Kiviat diagrams perhaps emphasizes the
shape of the tube and possibly conceals information about individ-
ual data items. These ideas are similar to our Parallel Glyphs in
that they link Star Glyphs but there is no integration with Parallel
Coordinates.
Finally, Wernert et al. [21] make an exploration of alternate multi-
dimensional data layouts but these representations are not linked
either visually, structurally, or interactively. (b) Parallel Coordinates unfolded in 3D.

3 PARALLEL G LYPHS : I NTEGRATING PARALLEL C OORDI - Figure 2: Advantage of unfolding Parallel Coordinates into 3D.
NATES WITH G LYPHS
As this test dataset illustrates, the problem of ambiguity created
Our technique integrates Parallel Coordinates with Star Glyphs to by overlapped polylines is likely to appear in real data, especially
create Parallel Glyphs. Parallel Coordinates are a 2D visual repre- when the number of data items increases.
sentation for multi-dimensional data (see Figure 2(a)) that uses a
set of vertical axes arranged left to right for plotting the values of 3.2 Construction of Parallel Coordinates in 3D
data items where each axis corresponds to one dimension. For each
data item, a polyline connects the points on each dimensional axis The Parallel Glyphs integrates Parallel Coordinates with Star
that correspond to the item’s value for that dimension. Glyphs by extending Parallel Coordinates into 3D space and unfold-
Star Glyphs are also a 2D visual representation for multi- ing them around a pivot axis (see Figure 3). In Parallel Coordinates
dimensional data (see Figure 1). Each glyph represents either a data data items are plotted as polylines over a set of dimension axes.

150
In order to be able to unfold the visualization into 3D we assign
each polyline representing a data item its own set of dimension axes
and utilize the baseline that connects the bottom of the dimension
axes as our pivot axis. Thus, while our 2D view of Parallel Glyphs
matches the original Parallel Coordinate view, without any further
transformations we can now successively apply rotations around the
pivot axis. This results in a visualization in which the data items
can be interactively unfolded into three dimensions. We can stop
or pause this opening at any chosen angle to change the degree of
unfolding, ranging from the original 2D view to a complete 360
degree rotation.
In the new visualization, Star Glyphs emerge during the unfold-
ing process. By connecting the outer end points of all line segments
for a given dimension the glyphs appear linked along the pivot axis
(see Figure 4). Thus, our 3D visualization naturally integrates both
types of visualizations for multidimensional data sets. Both Parallel
Coordinates and Star Glyphs can be shown connected to each other
and can be explored interactively.
When building the 3D visualization we considered both ortho-
graphic and perspective projection. The perspective projection is
typically used in applications that convey the sense of depth in a
more realistic way. With this projection, the length of the glyphs’
spikes varies, closer segments being shown longer than the farther
ones. These variations make it difficult for users to accurately visu-
ally compare the lengths of the spikes. Therefore, we decided to use
orthographic projection in our visual representation thus preserving
the lengths of the segments.
Traditionally, Parallel Coordinates have one axis for each dimen-
sion of the data set. Consequently, the Star Glyphs emerging in our
visualization have one glyph per dimension, each spike correspond-
ing to one data item. This is not the conventional way to map the
data in Star Glyphs: most related literature shows each glyph rep-
resenting all dimensions of a single item. In fact, our visualization
can generate this representation as well (see Figure 5) by merely
switching the way the data is read from the data table: changing
from columns over rows to rows over columns or vice versa. This
also results in a modified Parallel Coordinates representation that
shows dimensions as polylines and objects as axes (see Figure 6).
A series of questions are raised by this switch. In many situa-
tions there are more items than dimensions which would lead to an
increased number of glyphs aligned along the pivot line (as seen
in Figure 5). This will decrease space between them, raising the
question of the maximum reasonable size of data sets. However, by
adding some 3D interaction techniques to the new visualization it
may be possible to increase the number of glyphs supported and
enhance flexibility. For example, the pivot line that connects the
glyphs is just an abstract element and not a preferred dimension or
a time axis. Therefore, even if the order of glyphs is initially the
same as the order of items in the data table, this can be modified to
better support each task.
A disadvantage of a traditional 2D Star Glyph representation is
that the loss of context can lead to difficulties when conducting com-
parisons across glyphs. New interactions with the glyphs on the
pivot axis can address these issues. The pivot axis naturally aligns
the glyphs (see Figure 4), however, rotating the pivot axis can pro-
vide either an end view for direct comparison (see Figure 7(a)) or
it can be rotated so that the data item of interest is displayed top-
most (see Figure 7(b)). The orthographic projection makes com-
parisons valid and maintaining the polyline connections provided
by the Parallel Coordinates can also aid comparison (see Figure 8).
Additional 3D interactions are described in Section 4.

3.3 Interacting with Colour for Readability

In Parallel Coordinate visualizations the polylines representing the Figure 3: Several steps of unfolding the 2D Parallel Coordinates into
data items are typically drawn with a uniform colour or with the 3D Parallel Glyphs visualization.

151
 Figure 4: Star Glyphs emerging in the new visualization. Figure 8: Setting glyphs into relation to each other by arranging
them on the pivot axis and by connecting them with polylines.

proximity-based shades that convey information about the hierar-

chical structure of the data [3]. For glyphs in general colour is an
important attribute that is used to represent data [8, 24] in addition
to the size, position, orientation. However, for Star Glyphs in partic-
ular the most common way of mapping data attributes is not through
the colour but through the spikes’ lengths, which represent the val-
ues of the corresponding data item. Thus, a Star Glyph’s shape is
its main visual cue. Unfortunately, comparing spikes becomes in-
creasingly difficult with growing distance between them, especially
Figure 5: Star Glyphs, each glyph representing one object as opposed if one wants to compare spikes across more than one glyph. In
to one dimension as shown, e. g., in Figure 4. 2D, the glyphs’ spikes are drawn radially and it can be difficult to
accurately compare the lengths of spikes set at wide angles apart
from each other. 3D rotations can create further difficulties. Or-
thographic projection does make these comparisons valid for views
that are orthogonal to any two of the x, y, and z-dimensions.
We address this problem by adding circular textures to the glyphs
as additional visual cue. We create these textures by adjusting uni-
form colour scales [10] (see Figure 9) to offer the best readability
for glyphs, according to the following three criteria:
• provide a means for comparing pixels situated at a distance,
Figure 6: Parallel Coordinate representation resulting from Figure 5. • follow some natural ordering of colours, and
• be readable by colour-blind people.
While uniform colour scales are discussed extensively in literature
[10], these scales are typically based on algorithms that attempt to
maximize the number of just noticeable differences (JND) between
neighbouring pixels. This conflicts with the first criterion above.
Instead of JND, we are interested in recognizable differences across
a distance.
The approach we take extends the notion of a uniform colour
scale. We use a set of stepped scales with 16, 10 and 8 steps based
on the numerical values provided by [11] (see Figure 10). Each
sector of a stepped colour scale has a single colour that is obtained
by averaging the amount of red, green, and blue of corresponding
pixels from the uniform colour scale. This approach allows the
(a) Side view. (b) Front view. colour scale to meet the second criterion above (the natural sense of
order induced by a uniform scale) while adding the ability to easily
Figure 7: Star Glyphs aligned to coordinate axes to allow comparison notice ranges of values (first criterion).
of data items. According to [10] the heated and magenta scales are recom-
mended because the human eye can better notice variations in lu-
minance along the yellow-orange axis and variations in hue along
the magenta axis. Knowing that users have different preferences

152
Figure 10: Stepped heated colour scales applied to the glyphs to increase comparability between glyphs using 8, 10, and 16 steps, respectively.

Figure 9: Uniform heated colour scale applied to the glyphs using Figure 11: Using rings as rulers to compare values between glyphs.
texture mapping. In this example it is now possible to tell that the big top spike on
the middle glyph is slightly shorter than the corresponding spike on
the glyph to its left and that it represents the second biggest value
when choosing the colour scale we provide several other options as for this dimension.
well. We also considered colour scales based on opposite colours
with the two opposite colours on each end of an axis, as recom-
mended in [20] but we chose not to use them for two important scales based on a stepped rainbow scale (see Figure 13) that has
reasons. First, the variation on the red-green channel would be diffi- a distinct colour for each step . With this colour scale it is easy to
cult for colour-blind people (contradicting the third criterion above). identify and select polylines having the values corresponding to the
Second, the variations on the blue-yellow channel are hard to distin- main dimension in the same range. This is valid even if the poly-
guish on very thin sectors [20] as is likely to happen in glyphs with lines are not adjacent or if the glyphs are folded up or only slightly
many spikes. rotated. Of course, the rainbow scale will not meet the second crite-
We complement the support offered by a uniform colour scale rion above (natural order of colours) and so a legend for the colour
with a ring, centred about the pivot point for a glyph, that can be scale as an aid to the visualization reader is recommended.
interactively sized (see Figure 11). When the user clicks a point
inside a glyph, this ring is drawn on each glyph as a reference for 4 I NTERACTION
comparison. The rings have a radius equal to the distance between
the reference point and the centre of its corresponding glyph. These Data exploration using Parallel Glyphs is enhanced by several in-
rings act like a ruler and were created to assist in fine comparisons. teractions, some of which are specific to either the 2D or the 3D
The colour scales can also be applied to the polylines, having visualization. We extend the common techniques possible with 2D
one main dimension selected as a criterion for choosing the colour Parallel Coordinates such as brushing the polylines and dimensions,
(see Figure 12). This is especially useful when using the 2D Par- filtering tuples or dimensions, and reordering dimensions with new
allel Coordinates visualization to observe, for example, data trends. interactions that originate from the 3D nature of our visualization.
However, when using a stepped colour scale based on a uniform It is now possible to unfold the Parallel Glyphs using various an-
one, the steps may not be sufficiently distinguishable for other tasks gles as described in Section 3.2 and apply 3D manipulations to the
such as selection. resulting visualization. For example, we are able to rotate the visu-
To mitigate this problem we provide a set of stepped colour alization in all three directions around the pivot axis. A particularly

153
 the same dimension are plotted along the same axis, their compar-
ison is not necessarily easy because they might be overlapped by
other polylines. Parallel Glyphs offer the best of both worlds, sup-
porting the interactive adjustment of overlapping polylines and the
possibility of placing items adjacently for a given task.
Besides facilitating the comparison of data items, the polyline
reordering can lead to identification of patterns or trends in the data
set, as dimension reordering has been already shown to be useful for
this type of task. Our technique provides interactive item reordering
as the user might have knowledge about the data that would elude
an algorithm. However, an automatic operation equivalent to the
automatic dimension reordering would be interesting to develop.
The fact that the angle between two consecutive spikes of the
glyphs is no longer constant opens up new paths for data explo-
ration. In Parallel Coordinates, in order to analyze a subset of the
data independently from the rest of the set, the items not important
for the task are either made invisible or faded out. In our technique,
the items involved in a task can be maintained in the top side of
Figure 12: Brush-selecting a range of polylines with the help of a
stepped colour scale applied to the lines based on a selected dimen-
the screen while the remaining subset can be rotated to the bottom.
sion. The colours have been assigned according to the middle dimen- Thus, the entire set is visible and available for selection at any mo-
sion. Note that even with stepping applied to the colour scale it is ment, giving the user full overview of the data and making it easy to
hard to make out differences. switch between items from one category of interest to another. Both
the top and bottom subsets of the data items benefit independently
from different angles of unfolding, while the entire model can be
rotated in 3D, in order to find the best view (see Figure 14).

Figure 13: Using a stepped rainbow scale helps to increase the con-
trast between adjacent colour ranges to better support brushing. In
addition, trends are more visible for the entire dataset compared to
Figure 12.

interesting situation is when the glyphs are unfolded and the entire
model is rotated such that the glyphs directly overlap each other.
Translucent rendering of glyph interiors supports direct comparison Figure 14: Rearrangement of selected data items in 3D for compari-
of all glyphs (see Figure 7(a)). In addition to these interactions, we son without loosing context information.
implemented several new techniques described in the following.

4.1 Tuple Rearrangement 4.2 Lens Techniques

Having transformed Parallel Coordinates to three dimensions, the Depending on the data set, the visual density of the representation
order in which tuples are arranged now also has a meaning in ad- can make it extremely difficult to perform visual selection of po-
dition to the order of dimensions. Thus, we provide the possibility lylines that satisfy a given criterion or to identify overlapped data
to interactively change this order. The user can select one (through items. We address this problem by integrating focus + context capa-
clicking) or several (through brushing) polyline(s) and then to rotate bilities using two types of lenses.
them around the pivot axis to a different position in the sequence. The first one, E DGE L ENS [22] facilitates the exploration of the
This allows one to adapt the sequence of polylines in the data set to data set in Parallel Coordinates by adjusting the segments that fall
directly compare polylines that were previously far apart. Thus, the into its area of influence into Beziér curves (see Figure 15(a)).
user can avoid one of the main issues of a Star Glyphs visualization, These curves are shuffled at gradually decreasing distances from
the difficult comparison of non-adjacent spikes. Although in Paral- the centre of the lens, thus expanding the space available between
lel Coordinates the values of different data items corresponding to polylines and providing more insight into the dataset. The centre

154
of the lens can be easily moved until the desired presentation is
obtained. The desired polylines can then be brush selected.
For situations where the desired data items have a particular
value or range of values for a specific dimension, a second type
of lens may be used. For ease of selection purposes the EPS LENS
produces a non-linear expansion of the axis being in its immediate
vicinity (see Figure 15(b)). Thus, the points plotted along this axis
are separated by increased distances, allowing the identification and
selection of polylines required by a particular visual task.

(a) Edge lens. (b) EPS lens.

Figure 15: Lens interaction with Parallel Coordinates.

Figure 16: Task 1: Comparison of a sequence of data values using

repositioned Star Glyphs.
5 C ASE S TUDY

As a concrete example, we apply Parallel Glyphs to a set of 100 left) has peaks where the other two have local minima. This means
generations of plants generated by a genetic algorithm, each having that in generations where the width was small the height was also
five attributes: number of flowers, the dimensions of the plant in 3D small but the depth was inversed. This display suggests that the
(width, depth, and height), and the fitness factor. These are the five plants in each generation varied their dimensions to try to maintain
dimensions used in the visual representation, with each generation a constant fitness and volume.
represented by one polyline. In this section we look at how the
integration of Parallel Coordinates and Star Glyphs along with the
available colour scales support the user in solving a series of tasks.
The first task is to compare the variation of flowers and fitnesses
along the 100 generations. In Parallel Coordinates it is straightfor-
ward to compare the range of values plotted along the first and last
axis. However, this representation does not provide information
about how these values vary from one generation to another. This
task is easily solved with our technique by unfolding the glyphs
and by bringing the two desired glyphs next to each other (see Fig-
ure 16). In combination with the interactive ring ruler and with ro-
tations in any of the three dimensions, the variation between these (a) Data items selected with help of (b) Unfolded selected set of data
two attributes is easily compared, revealing a dependence relation- colour scale. items.
ship between the number of flowers and the fitness factor (the two
Star Glyphs in front in Figure 16). Figure 17: Task 2: examination of a selected subset of the data.
In the second task, one wants to understand how the shape of
plant changes over a selected number of generations for the plants
that have a fitness factor of between 70%–80%. First, to visually
select the desired items, the user applies the rainbow scale with 6 C ONCLUSION AND F UTURE W ORK
10 steps to the whole set of polylines, using the values of last di-
mension as a criterion for choosing the colour. Thus, each distinct In this paper we have introduced Parallel Glyphs which are a three-
colour of the colour scale corresponds to a range of 10% of the max- dimensional extension of Parallel Coordinates integrated with Star
imum fitness. Next, polylines that have a colour corresponding to Glyphs. Our work demonstrates that Parallel Coordinates and Star
the third sector of the colour scale (green) must be selected, but be- Glyphs are not only closely related and in the same family of visu-
cause the lines are too close to each other, accurate selection is very alizations but that they are, in fact, orthogonal to each other. We
difficult. Using the EPS LENS the user expands the space between showed that the traditional view of Parallel Coordinates—objects
those lines and then brushes only the green ones (see Figure 17(a)). over dimensions—leads to one of two versions of Star Glyphs—
The unselected polylines can be either made invisible (as in the Fig- one glyph representing all values for one dimension. The other
ure) or rotated to the bottom of the screen to keep them available version of Star Glyphs—one glyph representing the values of one
for further tasks. By choosing to keep only the selected data items data object—can be derived from a second version of Parallel Coor-
the user is free to completely unfold the glyphs formed with the dinates in which dimensions are plotted over objects.
remaining items (see Figure 17(b)). The glyphs corresponding to We have enriched the new 3D visualization by applying colour
width and height (second and third from left) have similar shapes scales as textures to the 3D Star Glyphs to support the task of com-
but different sizes. This means the variation of these two attributes paring spatially separated glyphs. We discussed both uniform and
is similar. However, the glyph corresponding to depth (fourth from stepped colour scales and compared their merits. We also showed

155
 how these colour scales can lead to colouring the data items’ po- [12] Allen R. Martin and Matthew O. Ward. High Dimensional Brushing
lylines in order to identify trends and support brush-selection. To for Interactive Exploration of Multivariate Data. In Proceedings of
further support visualization tasks we discussed various interaction IEEE Visualization 1995 (VIS 1995), pages 271–278, Los Alamitos,
possibilities. In addition to well-known techniques for 2D Parallel CA, 1995. IEEE Computer Society Press.
Coordinates, the unfolding of 2D Parallel Coordinates into the new [13] Wei Peng, Matthew O. Ward, and Elke A. Rundensteiner. Clutter
3D representation, and 3D transformations such as rotation around Reduction in Multi-Dimensional Data Visualization Using Dimension
coordinate axes we introduced a ring ruler interaction that enables Reordering. In Proceedings of the IEEE Symposium on Information
determining even small differences between Star Glyphs. Further- Visualization 2004 (InfoVis 2004), pages 89–96, Los Alamitos, CA,
more, we discussed the newly possible data item rearrangement 2004. IEEE Computer Society Press.
[14] William Ribarsky, Eric Ayers, John Eble, and Sougata Mukherjea.
that enables users to directly compare selected data items with each
Glyphmaker: Creating Customized Visualizations of Complex Data.
other without loosing the context of the rest of the data set. Finally,
Computer, 27(7):57–64, July 1994.
we discussed enhancements to interactions that are applied to the [15] Christopher D. Shaw, James A. Hall, Christine Blahut, David S. Erbert,
traditional 2D Parallel Coordinates visualization such as lens tech- and D. Aaron Roberts. Using Shape to Visualize Multivariate Data. In
niques for focus + context data exploration. Proceedings of the 1999 Workshop on New Paradigms in Information
In the future we will examine the limits for the supported size Visualization and Manipulation, pages 17–20, New York, 1999. ACM
of data set for the technique to be usable. In addition, we envision Press.
adding shading and other 3D rendering techniques to improve the [16] Harri Siirtola. Direct Manipulation of Parallel Coordinates. In Pro-
visualization technique. Also, we will explore 3D shape analysis ceedings of IEEE Visualization 2000 (VIS 2000), pages 373–378, Los
methods to extract more information from the data set and research Alamitos, CA, 2000. IEEE Computer Society Press.
automatic reordering of data items. [17] Christian Tominski, James Abello, and Heidrun Schumann. Axes-
Based Visualizations with Radial Layouts. In Proceedings of the 2004
ACM Symposium on Applied Computing, pages 1242–1247, New
ACKNOWLEDGEMENTS York, 2004. ACM Press.
[18] Matthew O. Ward. XmdvTool: Integrating Multiple Methods for Visu-
We gratefully thank our funding provider Natural Sciences and En- alizing Multivariate Data. In Proceedings of IEEE Visualization 1994
gineering Research Council of Canada (NSERC). We would also (VIS 1994), pages 326–333, Los Alamitos, CA, 1994. IEEE Computer
like to thank researchers from the Interactions Lab at the Univer- Society Press.
sity of Calgary for their insightful comments on this work and, in [19] Matthew O. Ward. Creating and Manipulating N-Dimensional
particular, Michael Boyle for his help in editing this paper. Brushes. In Proceedings of Joint Statistical Meeting 1997, pages 6–14,
1997.
R EFERENCES [20] Colin Ware. Information Visualization. Morgan Kaufmann Publishers,
San Francisco, 2000.
[1] Almir Olivette Artero, Maria Cristina Ferreira de Oliveira, and Haim [21] Eric A. Wernert, Donald K. Berry, John N. Huffman, and Craig A.
Levkowitz. Uncovering Clusters in Crowded Parallel Coordinates Vi- Stewart. Tree3d-A System for Temporal and Comparative Analysis of
sualizations. In Proceedings of the IEEE Symposium on Information Phylogenetic Trees. In Proceedings of the IEEE InfoVis 2003 Poster
Visualization 2004 (InfoVis 2004), pages 81–88, Los Alamitos, CA, Compendium, pages 114–115, Los Alamitos, CA, 2003. IEEE Com-
2004. IEEE Computer Society Press. puter Society Press.
[2] Hong Chen. Compound Brushing. In Proceedings of the IEEE Sym- [22] Nelson Wong, Sheelagh Carpendale, and Saul Greenberg. EdgeLens:
posium on Information Visualization 2003 (InfoVis 2003), pages 181– An Interactive Method for Managing Edge Congestion in Graphs. In
188, Los Alamitos, CA, 2003. IEEE Computer Society Press. Proceedings of the IEEE Symposium on Information Visualization
[3] Ying-Huey Fua, Matthew O. Ward, and Elke A. Rundensteiner. Hi- 2003 (InfoVis 2003), pages 51–58, Los Alamitos, CA, 2003. IEEE
erarchical Parallel Coordinates for Exploration of Large Datasets. In Computer Society Press.
Proceedings of IEEE Visualization 1999 (VIS 1999), pages 43–50, Los [23] Pak Chung Wong and R. Daniel Bergeron. Multivariate Visualization
Alamitos, CA, 1999. IEEE Computer Society Press. Using Metric Scaling. In Proceedings of IEEE Visualization 1997
[4] Steven T. Hackstadt and Allen D. Malony. Visualizing Parallel Pro- (VIS 1997), pages 111–118, Los Alamitos, CA, 1997. IEEE Computer
grams and Performance. IEEE Computer Graphics and Applications, Society Press.
15(4):12–14, July 1995. [24] Jing Yang, Anilkumar Patro, Shiping Huang, Nishant Mehta,
[5] Helwig Hauser, Florian Ledermann, and Helmut Doleisch. Angular Matthew O. Ward, and Elke A. Rundensteiner. Value and Relation Dis-
Brushing of Extended Parallel Coordinates. In Proceedings of the play for Interactive Exploration of High Dimensional Datasets. In Pro-
IEEE Symposium on Information Visualization 2002 (InfoVis 2002), ceedings of the IEEE Symposium on Information Visualization 2004
pages 127–130, Los Alamitos, CA, 2002. IEEE Computer Society (InfoVis 2004), pages 73–80, Los Alamitos, CA, 2004. IEEE Com-
Press. puter Society Press.
[6] Alfred Inselberg and Bernard Dimsdale. Parallel Coordinates: A Tool [25] Jing Yang, Wei Peng, Matthew O. Ward, and Elke A. Rundensteiner.
for Visualizing Multi-Dimensional Geometry. In Proceedings of IEEE Interactive Hierarchical Dimension Ordering, Spacing and Filtering
Visualization 1990 (VIS 1990), pages 361–378, Los Alamitos, CA, for Exploration of High Dimensional Datasets. In Proceedings of the
1990. IEEE Computer Society Press. IEEE Symposium on Information Visualization 2003 (InfoVis 2003),
[7] Alfred Inselberg. The Plane with Parallel Coordinates. The Visual pages 105–112, Los Alamitos, CA, 2003. IEEE Computer Society
Computer, 1(2):69–91, October 1985. Press.
[8] Daniel A. Keim. Pixel-Oriented Database Visualizations. ACM SIG- [26] Jing Yang, Matthew O. Ward, Elke A. Rundensteiner, and S. Huang.
MOD Record, 25(4):35–39, December 1996. Visual Hierarchical Dimension Reduction for Exploration of High Di-
[9] Daniel A. Keim. Information Visualization and Visual Data Mining. mensional Datasets. In Proceedings of the ACM Symposium on Data
IEEE Transactions on Visualization and Computer Graphics, 8(1):1– Visualisation 2003, pages 19–28, 282, New York, 2003. ACM Press.
8, January 2002. [27] Jing Yang, Matthew O. Ward, and Elke A. Rundensteiner. InterRing:
[10] Haim Levkowitz and Gabor T. Herman. Color Scales for Image Data. An Interactive Tool for Visually Navigating and Manipulating Hierar-
IEEE Computer Graphics and Applications, 12(1):72–80, January chical Structures. In Proceedings of the IEEE Symposium on Infor-
1992. mation Visualization (InfoVis 2002), pages 77–84, Los Alamitos, CA,
[11] Haim Levkowitz. ColorCenterColormaps. Web site at http://www. 2002. IEEE Computer Society Press.
cs.uml.edu/∼haim/ColorCenter/ColorCenterColormaps.
htm. visited April 19, 2005.

156