Principles of Data Visualization

Eamonn Maguire, DPhil
Principles of Data Visualization
iCSC, CERN
March 2017

The role of visualization systems is to provide visual representations of datasets that
help people carry out tasks more effectively.
Visualization
Tamara Munzner
A Visualization should
1. Save time

2. Have a clear purpose*
3. Include only the relevant content*

4. Encodes data/information appropriately
* from Noel Illinsky, http://complexdiagrams.com/

Visualization is suitable when there is a need to augment human capabilities
rather than replace people with computational decision-making methods.
Visualization
Tamara Munzner
A Visualization should
1. Save time

2. Have a clear purpose*
3. Include only the relevant content*

4. Encodes data/information appropriately
* from Noel Illinsky, http://complexdiagrams.com/

Course outcomes
The what? Major data types and classiﬁcations of them
The why? Why are we visualising at all?
The how? How can we visualize? What archetypes can we use to guide us?
Finally: case study? Given some data, how can we go about visualising it?

A lot of the content for this introduction comes from this book from Prof.
Tamara Munzner (UBC, Vancouver, Canada) which I created the illustrations
for.
If you’re interested in learning more, it’s a great book to check out from the
CERN library, or buy :)

External representation:
replace cognition with
perception
Visualization

Cerebral:Visualizing Multiple Experimental Conditions on a
Graph with Biological Context. Barsky, Munzner, Gardy, and
Kincaid. IEEETVCG (Proc. InfoVis) 14(6):1253-1260,
2008.]
External representation:
replace cognition with
perception
Visualization

The statistics would lead us to
believing that everything is the
same
Why visualize?

A Nested Model of Visualization Design and Validation.
Munzner. IEEE TVCG 15(6):921-928, 2009 (Proc. InfoVis
2009).
Analysis framework: Four levels, three questions
Data/task abstraction
Visual encoding/interaction idiom
Algorithm
Domain situation

• Domain situation
2009).
Algorithm
Domain situation

– who are the target users?
2009).
Algorithm
Domain situation

• Data/Task Abstraction
2009).
Algorithm
Domain situation

– translate from speciﬁcs of domain to
vocabulary of vis
2009).
Algorithm
Domain situation

vocabulary of vis
• What is shown? Data abstraction
2009).
A Multi-Level Typology of Abstract Visualization Tasks
Brehmer and Munzner. IEEE TVCG 19(12):2376-2385,
2013 (Proc. InfoVis 2013).
Algorithm
Domain situation

vocabulary of vis
• Why is the user looking at it? Task
abstraction
2009).
Algorithm
Domain situation

vocabulary of vis
abstraction
• Visual Encoding
2009).
Algorithm
Domain situation

vocabulary of vis
abstraction
• Visual Encoding
• How is it shown?
2009).
Algorithm
Domain situation

vocabulary of vis
abstraction
• Visual Encoding
• visual encoding: how to draw
2009).
Algorithm
Domain situation

vocabulary of vis
abstraction
• Visual Encoding
• interaction: how to manipulate
2009).
Algorithm
Domain situation

vocabulary of vis
abstraction
• Visual Encoding
• Algorithm
2009).
Algorithm
Domain situation

vocabulary of vis
abstraction
• Visual Encoding
• Algorithm
– efﬁcient computation, layout
algorithms etc.
2009).
Algorithm
Domain situation

The branches of data visualization
Information Visualization Scientiﬁc Visualization
Position is given.
Also medical visualizations
Position is derived.
Incl. GeoVis

Discover
Finding new insights in your data
Implies a level of interactivity to query, compare, correlate etc.

Discover
Finding new insights in your data
This is typically where one should be careful in how information is presented.
An erroneous data encoding could bring about wrong conclusions.
Implies a level of interactivity to query, compare, correlate etc.

Present
Presenting your results, e.g. for a paper

Enjoy
Infographics, art, or superﬂuous visualizations.

Enjoy
Infographics, art, or superﬂuous visualizations.
Bear in mind that this was a non-interactive ﬁgure in a paper

How can you encode information optimally?

JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC
2 10 4 5 6 9 1 3 5 3 4 7

2 10 4 5 6 9 1 3 5 3 4 7
0
5
10
Scatter

2 10 4 5 6 9 1 3 5 3 4 7
0
5
10
0
5
10
Scatter Line

2 10 4 5 6 9 1 3 5 3 4 7
0
5
10
0
5
10
0
5
10
Scatter Line
Histogram

2 10 4 5 6 9 1 3 5 3 4 7
0
5
10
0
5
10
0
5
10
Scatter Line
Histogram Area
0
5
10

2 10 4 5 6 9 1 3 5 3 4 7
0
5
10
0
5
10
0
5
10
Scatter Line
Histogram Area
Size
0
5
10

2 10 4 5 6 9 1 3 5 3 4 7
0
5
10
0
5
10
0
5
10
JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC
Scatter Line
Histogram Area
Size Saturation
0
5
10

2 10 4 5 6 9 1 3 5 3 4 7
0
5
10
0
5
10
0
5
10
Scatter Line
Histogram Area
Size Saturation
Size & Saturation
0
5
10

2 10 4 5 6 9 1 3 5 3 4 7
0
5
10
0
5
10
0
5
10
Scatter Line
Histogram Area
Size Saturation
Size & Saturation
0
5
10
Size, Saturation, & Position
5

And that’s just a really simple low dimensional example
Moreover, all of these visualizations encode the information,
but the decode error (interpreting, comparing, …) for each is different
So, why?

Stevens, 1975
Our perception system does not behave linearly.
Some stimuli are perceived less or more than intended.

We have to be careful when mapping
data to the visual world
Some visual channels are more effective for some
data types over others.

Some data has a natural mapping that our
brains expect given certain types of data

There are many intricacies of the visual system that must be considered

The pop-out effect
• Parallel processing on many individual channels
– speed independent of distractor count
– speed depends on channel and amount of difference from
distractors
• Serial search for (almost all) combinations
– speed depends on number of distractors
We pre-attentively process a scene, and some visual elements
stand out more than others.

Not all exhibit the pop-out effect!

Parallel line pairs do not pop out from tilted pairs…

Parallel line pairs do not pop out from tilted pairs…
And not all visual channels pop out as quickly as other. E.g. colour is always on top.

Relative Comparison
4 values Unordered Unaligned

Relative Comparison

4 values
UnorderedAligned
Relative Comparison

4 values
UnorderedAligned
8 values
UnorderedAligned
Relative Comparison

8 values
20 values
Relative Comparison

Known Target Search
Colour Variety
Unknown Target Search
Grouped
Random
Grouped
Random
Random Grouped
Target shown before hand (known) or not shown (unknown).
The unique colour here is the orange square.
A) Known and Unknown Target Search
C) Response Time and Accuracy Results
Known Target Search
Colour Variety
Unknown Target Search
Grouped
Random
Grouped
Random
Subitizing
Grouped
Random
Random Grouped
Which grid has more colours?
7 8
B) Subitizing (how many colours?)
How Capacity Limits of Attention Inﬂuence Information Visualization Effectiveness.
Haroz S. and Whitney D., IEEE TVCG 2012

A. Law of Closure B. Law of Similarity D. Law of Connectedness E. Law of Symmetry
[ ] { } ( )
C. Law of Proximity
F. Law of Good
Continuation
G. Contour Saliency
a c
b
a
c
b
d
H. Law of Common Fate I. Law of Past
Experience
J. Law of
Pragnanz
K. Figure/Ground
Gestalt Laws

Dimension X
A and B have the same width.
However B and C are perceived
more alike even though they are
different widths and heights.
Width and Height are integral
dimensions
Dimension X
Width
Integral/Separable Dimensions

Dimension X
A and B have the same width.
However B and C are perceived
more alike even though they are
different widths and heights.
Width and Height are integral
dimensions
Dimension X
Width
Dimension X
Colour
A and B have the same
colour and are perceived
more similar.
Colour and Height are
Separable dimensions

Fully separableFully Integral
Dimension X
Width
Dimension X
Orientation
Dimension X
Colour
Dimension X
Motion
Dimension X
Motion
Dimension X
Colour

We don’t see in 3D, and we have difﬁculties interpreting
information on the Z-axis.
There are many visual tricks that can be observed due to
how the visual system works

2D always wins…
Our visual system is not good at interpreting information on
the z-axis.
*3D is normally only used for exploration of inherently 3D information, such as
medical imaging data…

These options, taken randomly from google image searches so how widely 3D is abused in
information visualization. All of these charts are manipulating our perception of the data by
using the Z axis to occlude information…it would be avoided in 2D.
2D always wins…

http://cms-results.web.cern.ch/cms-results/public-results/preliminary-results/BPH-14-008/index.html
2D always wins…

We don’t see in 3D, and we have difﬁculties interpreting
information on the Z-axis.
There are many visual tricks that can be observed due to
how the visual system works
Colour

The simplest, yet most abused of all visual encodings.
The problem is that a smooth step in a value does not equate to a
smooth colour transition…
Colour

Additionally, colour is not equally binned in reality. We perceive
colours differently due to an increased sensitivity to the yellow part
of the spectrum…
Wavelength (nm)
IRUV
Visible Spectrum

https://mycarta.wordpress.com/2012/10/06/the-rainbow-is-deadlong-live-the-rainbow-part-3/
Luminosity is also not stable across the colours, meaning some colours
will pop out more than others… and not always intentionally.

Gregory compared the wavelength of light with the smallest observable difference
in hue (expressed as wavelength difference)
And how we perceive changes in hue is also very different.

Is this a good visualization?
But, grayscale would be just as useful and less visually distracting.

Is there a colour palette for scientiﬁc
visualization that works?

HSL linear L rainbow palette
Kindlmann, G. Reinhard, E. and Creem, S., 2002, Face-based Luminance Matching for Perceptual Colormap
Generation, IEEE Proceedings of the conference on Visualization ’02
https://mycarta.wordpress.com/2012/10/06/the-rainbow-is-deadlong-live-the-rainbow-part-3/

But there are some in CMS that are already moving
away from the rainbow
http://cms-results.web.cern.ch/cms-results/public-results/publications/
JME-13-004/index.html

Is there a colour palette for scientiﬁc visualization
that works?

Binary
Diverging
Categorical
Sequential
Categorical
Categorical
There are also lots of default colour maps that can be applied to
particular data types.
http://colorbrewer2.org/

Semantic relevance
Or just consistency
When there are many colours for example, we ﬁnd it
difﬁcult to remember abstract associations.

Selecting Semantically-Resonant Colors for Data Visualization
Sharon Lin, Julie Fortuna, Chinmay Kulkarni, Maureen Stone, Jeffrey Heer
Computer Graphics Forum (Proc. EuroVis), 2013
What are semantically resonant colours?

Semantic colouring is a good idea in theory, but
there are limited areas where this really works.
But, if you are going to use colour, have a consistent
colour mapping. That way, the decoding time is less.
Saving time…
In general, CMS & other experiments are pretty good at
this, but there are some examples…

http://cms-results.web.cern.ch/cms-results/public-
results/preliminary-results/SUS-16-029/CMS-PAS-
SUS-16-029_Figure_003.png

What happens when we have a high number of
dimensions?
And that was just to represent a low number
of dimensions

Parallel
Coordinates
Scatter Plot
Matrices
Glyphs
Linked
Plots
Height Weight Cholesterol
Multidimensional Visualization

Scatter Plot Matrices
…
Height Weight CholName
1.76 63 4.5John
1.79 70 4.15Mike
1.61 60 6.7Jim
1.84 90 5.03Francois

…
Height Weight CholesterolHeight Weight CholName
1.76 63 4.5John
1.79 70 4.15Mike
1.61 60 6.7Jim

…
1.76 63 4.5John
1.79 70 4.15Mike
1.61 60 6.7Jim
1.76

…
1.76 63 4.5John
1.79 70 4.15Mike
1.61 60 6.7Jim
1.76
63
1.76
4.5

…
1.76 63 4.5John
1.79 70 4.15Mike
1.61 60 6.7Jim
1.76
63
1.76
4.5
Height Weight Cholesterol

Visual Exploration of Large Structured Datasets. Wills. Proc. New Techniques and Trends in
Statistics (NTTS), pp. 237–246. IOS Press, 1995.
Linked Plots

When one visualization won’t cut it…

With dc.js, crossﬁlter, and d3.js

My Tutorial on Creating Dashboard Visualizations
https://thor-project.github.io/dashboard-tutorial/

Parallel Coordinate Plots
Positive Correlation Negative (inverse) Correlation No Correlation

Lets take an example where we have many variables to display...
Each user is represented by a circle
user a
user z
user a
user z
2 Dimensions 3 Dimensions
Size indicates number of logins per day

As we get to higher levels of dimensions, we’ll have problems. Our choice of visual encoding will affect the visual
availability of each dimension to the user.
4 Dimensions
Color indicates users department
5 Dimensions
Transparency indicates consistency in logins
user a
user z
user a
user z

Parallel coordinates are a visualization technique employed when a large number of dimensions need to be displayed
(often without a temporal element) and where each of those dimensions can be equally important in the decision
making process.
Not so easy to spotEasy to see
In the scatter plots here, it’s easy to see correlation between downloads and uploads,
but with the other dimensions that’s difficult.

Positive Correlation Negative (inverse) Correlation No Correlation

2 Dimensions
Uploads
Downloads

2 Dimensions
Uploads
Downloads
1 User

2 Variables
Uploads
Downloads
11 Users

2 Dimensions
Uploads
Downloads
11 Users
user a
user z

3 Dimensions
Uploads
Downloads
Logins per day
11 Users

4 Dimensions
Uploads
Downloads
Logins per day
Std. deviation in logins per day
11 Users
2
-2

5 Dimensions
Uploads
Downloads
Logins per day
Std. deviation in logins per day
Department
11 Users We use color for department since it’s categorical information.
2
-2
We can keep adding more parallel lines, and comfortably have
around 20 dimensions for many users displayed at once.

Clustering Outlier DetectionCorrelation Distribution
Parallel coordinates provide an eﬃcient way to visualize many variables, along with
their associated clusters, anomalies, value distributions and correlations.

83
• static item aggregation
• task: ﬁnd distribution
• data: table
• derived data
– 4 quantitative attributes
• median: central line
• lower and upper quartile:
boxes
• lower upper fences:
whiskers
– outliers beyond fence cutoffs
explicitly shown
and kurtosis 20) and a bimo
tion 0.31). Richer displays o
multi-modality is particularly
!
!
!!
!
!
!
!
!
n s k mm!2024
!2024
Figure 4: From left to right: box
Glyphs

Glyphs

A Simple Example | Student Test Results
Math
Physics
English
Religion
Math Physics English Religion
Table Scatter Plot Matrix
85
90
65
50
40
95
80
50
40
60
71
60
90
95
80
65
50
90
80
90

86
85
90
65
50
40
95
80
50
40
60
71
60
90
95
80
65
50
90
80
90
Table Parallel Coordinates
100
90
80
70
60
50
40
30
20
10
0

87
85
90
65
50
40
95
80
50
40
60
71
60
90
95
80
65
50
90
80
90
Table Glyph
Math
Physics
English
Religion

87
85
90
65
50
40
95
80
50
40
60
71
60
90
95
80
65
50
90
80
90
Table Glyph
Math
Physics
English
Religion
Teacher
Arrange Spatially

Graphs/Networks
Cerebral:Visualizing Multiple Experimental Conditions on a Graph with Biological Context. Barsky, Munzner,
Gardy, and Kincaid. IEEETVCG (Proc. InfoVis) 14(6):1253-1260, 2008.]

Graphs/Networks
https://cambridge-intelligence.com

Graphs/Networks
Matrix Representations
https://bost.ocks.org/mike/miserables/

Graphs/Networks
Hive Plots
http://jsﬁddle.net/7a7b5dwp/

Graphs/Networks
Hive Plots
http://jsﬁddle.net/7a7b5dwp/ http://jsﬁddle.net/eamonnmag/vso70qnr/

Trees
Cut Level 0.3 4 Clusters
View as treemapDendrogram Search

• split by
neighbourhood
• then by type
• colour by price
• neighbourhood
patterns
– where it’s expensive
– where you pay much
more for detached
type
94
Conﬁguring Hierarchical Layouts to Address Research Questions. Slingsby, Dykes, andWood. IEEETransactions on
Visualization and Computer Graphics (Proc. InfoVis 2009) 15:6 (2009), 977–984.
Treemaps Partitioning

• switch order of splits
– type then
neighbourhood
• switch colour
– by price variation
• type patterns
– within speciﬁc
type, which
neighbourhoods
are inconsistent
95
Conﬁguring Hierarchical Layouts to Address Research Questions. Slingsby, Dykes, andWood. IEEETransactions on
Visualization and Computer Graphics (Proc. InfoVis 2009) 15:6 (2009), 977–984.
Treemaps Partitioning

97
Validating your visualisation…

97
Validating your visualisation…
Domain situation
Observe target users using existing tools
Justify design with respect to alternatives
Algorithm
Measure system time/memory
Analyze computational complexity
Observe target users after deployment ( )
Measure adoption
Analyze results qualitatively
Measure human time with lab experiment (lab study)

98
“Visualization can surprise you, but doesn't scale well.
Modelling scales well, but can't surprise you.”
Hadley Wickham
The elephant in the room… scaling up

99
One of the biggest challenges in HEP, Biology, chemistry,
and business is scale.
Our screens have a limited number of pixels. And our data
is often much larger.
You can do two things to get around this.
The elephant in the room… scaling up

100
GPU usage and
WebGL
Not yet compatible across all
browsers, and not everyone
has a dedicated GPU.
Reduce the
problem space
Try and get users to focus
queries as soon as possible
to reduce the amount of
data to be visualised.
As done in imMens from
Jeffrey Heer’s lab in
Washington. Renders billions
of data points at 50 fps in
the browser.
Provide ways to aggregate
information in to meaningful
overview visualisations…
Solutions

101
Munzner. A K Peters Visualization Series, CRC Press, Visualization Series, 2014.
Visualization Analysis and Design.
More??

102
http://antarctic-design.co.uk/biovis-workshop15/
Tutorial on D3
Tutorial on Dashboard Visualizations
https://thor-project.github.io/dashboard-tutorial/
Visualization Survey Sites
Set Visualization - http://www.cvast.tuwien.ac.at/SetViz
Time Series Visualization - http://survey.timeviz.net/
Parallel Coordinates Visualization
Periodic Table of Visualizations
Data Vis Catalogue
Further Links

Eamonn Maguire
CERN
Data Visualization and suggestions for CMS
CERN
August 25th 2016
Questions
@antarcticdesign
eamonnmag@gmail.com

Principles of Data Visualization

More Related Content

What's hot

Similar to Principles of Data Visualization

More from Eamonn Maguire

Recently uploaded

In this document

Principles of Data Visualization