DSC 201: Data Analysis & Visualization · 2018. 10. 4. · 4 6 8 10 12 14 16 18 4 6 8 10 12 x 1 y 1...
Transcript of DSC 201: Data Analysis & Visualization · 2018. 10. 4. · 4 6 8 10 12 14 16 18 4 6 8 10 12 x 1 y 1...
DSC 201: Data Analysis & Visualization
Visualization Dr. David Koop
D. Koop, DSC 201, Fall 2018
“Computer-based visualization systems provide visual representations of datasets designed to help people carry out tasks more effectively.”
— T. Munzner
�2D. Koop, DSC 201, Fall 2018
Why Visual?
�3D. Koop, DSC 201, Fall 2018
[F. J. Anscombe]
I II III IV
x y x y x y x y
10.0 8.04 10.0 9.14 10.0 7.46 8.0 6.58
8.0 6.95 8.0 8.14 8.0 6.77 8.0 5.76
13.0 7.58 13.0 8.74 13.0 12.74 8.0 7.71
9.0 8.81 9.0 8.77 9.0 7.11 8.0 8.84
11.0 8.33 11.0 9.26 11.0 7.81 8.0 8.47
14.0 9.96 14.0 8.10 14.0 8.84 8.0 7.04
6.0 7.24 6.0 6.13 6.0 6.08 8.0 5.25
4.0 4.26 4.0 3.10 4.0 5.39 19.0 12.50
12.0 10.84 12.0 9.13 12.0 8.15 8.0 5.56
7.0 4.82 7.0 7.26 7.0 6.42 8.0 7.91
5.0 5.68 5.0 4.74 5.0 5.73 8.0 6.89
Why Visual?
�3D. Koop, DSC 201, Fall 2018
[F. J. Anscombe]
I II III IV
x y x y x y x y
10.0 8.04 10.0 9.14 10.0 7.46 8.0 6.58
8.0 6.95 8.0 8.14 8.0 6.77 8.0 5.76
13.0 7.58 13.0 8.74 13.0 12.74 8.0 7.71
9.0 8.81 9.0 8.77 9.0 7.11 8.0 8.84
11.0 8.33 11.0 9.26 11.0 7.81 8.0 8.47
14.0 9.96 14.0 8.10 14.0 8.84 8.0 7.04
6.0 7.24 6.0 6.13 6.0 6.08 8.0 5.25
4.0 4.26 4.0 3.10 4.0 5.39 19.0 12.50
12.0 10.84 12.0 9.13 12.0 8.15 8.0 5.56
7.0 4.82 7.0 7.26 7.0 6.42 8.0 7.91
5.0 5.68 5.0 4.74 5.0 5.73 8.0 6.89
Mean of x 9Variance of x 11Mean of y 7.50Variance of y 4.122Correlation 0.816
●
●●
●●
●
●
●
●
●●
4 6 8 10 12 14 16 18
4
6
8
10
12
x1
y 1
●●
●●●
●
●
●
●
●
●
4 6 8 10 12 14 16 18
4
6
8
10
12
x2
y 2●
●
●
●●
●
●●
●
●●
4 6 8 10 12 14 16 18
4
6
8
10
12
x3
y 3
●●
●
●●
●
●
●
●
●
●
4 6 8 10 12 14 16 18
4
6
8
10
12
x4
y 4
Why Visual?
�4D. Koop, DSC 201, Fall 2018
[F. J. Anscombe]
Designing Visualizations
�5D. Koop, DSC 201, Fall 2018
[M. Stefaner, 2013]
Visualization Effectiveness
�6D. Koop, DSC 201, Fall 2018
[A. Kitaoka]
Visualization Effectiveness
�6D. Koop, DSC 201, Fall 2018
[A. Kitaoka]
Red, yellow, blue
Purple, orange do not exist!
Channels: Visual Appearance• How should we encode this data?
�7D. Koop, DSC 201, Fall 2018
Name Region Population Life Expectancy Income
China East Asia & Pacific 1335029250 73.28 7226.07
India South Asia 1140340245 64.01 2731
United States America 306509345 79.43 41256.08
Indonesia East Asia & Pacific 228721000 71.17 3818.08
Brazil America 193806549 72.68 9569.78
Pakistan South Asia 176191165 66.84 2603
Bangladesh South Asia 156645463 66.56 1492
Nigeria Sub-Saharan Africa 141535316 48.17 2158.98
Japan East Asia & Pacific 127383472 82.98 29680.68
Mexico America 111209909 76.47 11250.37
Philippines East Asia & Pacific 94285619 72.1 3203.97
Vietnam East Asia & Pacific 86970762 74.7 2679.34
Germany Europe & Central Asia 82338100 80.08 31191.15
Ethiopia Sub-Saharan Africa 79996293 55.69 812.16
Turkey Europe & Central Asia 72626967 72.06 8040.78
Share ! " #Bubbles $
Color
Select
Size
Zoom20152015
30
40
50
60
70
80
year
s
Life
exp
ecta
ncy ▼
1800 1900 2000
World Regions
Search...
Afghanistan
Albania
Algeria
Andorra
Angola
Antigua and Barbuda
Argentina
Armenia
Australia
Austria
Azerbaijan
Bahamas
Bahrain
Bangladesh
Barbados
Belarus
Population, total
100100%%
OPTIONS EXPAND PRESENT
English ▼ FACTS TEACH ABOUT ►HOW TO USE
Potential Solution
�8D. Koop, DSC 201, Fall 2018
[Gapminder, Wealth & Health of Nations]
Assignment 2• Visualizing Hurricane Data using Tableau and Altair • Tasks:
- Statistics - Number of hurricanes over the years - Windspeed versus pressure - Locations of hurricanes
�9D. Koop, DSC 201, Fall 2018
Visual Encoding• How do we encode data visually?
- Marks are the basic graphical elements in a visualization - Channels are ways to control the appearance of the marks
• Marks classified by dimensionality:
• Also can have surfaces, volumes • Think of marks as a mathematical definition, or if familiar with tools
like Adobe Illustrator or Inkscape, the path & point definitions
�10D. Koop, DSC 201, Fall 2018
Points Lines Areas
Horizontal
Position
Vertical Both
Color
Shape Tilt
Size
Length Area Volume
Visual Channels
�11D. Koop, DSC 201, Fall 2018
[Munzner (ill. Maguire), 2014]
Channels• Usually map an attribute to a single channel
- Could use multiple channels but… - Limited number of channels
• Restrictions on size and shape - Points are nothing but location so size and shape are ok - Lines have a length, cannot easily encode attribute as length - Maps with boundaries have area, changing size can be
problematic
�12D. Koop, DSC 201, Fall 2018
Cartograms
�13D. Koop, DSC 201, Fall 2018
[Election Results by Population, M. Newman, 2012]
Channel Types• Identity => what or where, Magnitude => how much
�14D. Koop, DSC 201, Fall 2018
[Munzner (ill. Maguire), 2014]
Magnitude Channels: Ordered Attributes Identity Channels: Categorical Attributes
Spatial region
Color hue
Motion
Shape
Position on common scale
Position on unaligned scale
Length (1D size)
Tilt/angle
Area (2D size)
Depth (3D position)
Color luminance
Color saturation
Curvature
Volume (3D size)
Channels: Expressiveness Types and Effectiveness Ranks
Mark Types• Can have marks for items and links
- Connection => pairwise relationship - Containment => hierarchical relationship
�15D. Koop, DSC 201, Fall 2018
[Munzner (ill. Maguire), 2014]
Marks as Items/Nodes
Marks as Links
Points Lines Areas
Containment Connection
Expressiveness and Effectiveness• Expressiveness Principle: all data from the dataset and nothing
more should be shown - Do encode ordered data in an ordered fashion - Don’t encode categorical data in a way that implies an ordering
• Effectiveness Principle: the most important attributes should be the most salient - Saliency: how noticeable something is - How do the channels we have discussed measure up? - How was this determined?
�16D. Koop, DSC 201, Fall 2018
How do we test effectiveness?
�17D. Koop, DSC 201, Fall 2018
esting set of perceptual tasks, we replicated Cleveland &McGill’s [7] classic study (Exp. 1A) of proportionality es-timates across spatial encodings (position, length, angle),and Stone & Bartram’s [30] alpha contrast experiment (Exp.2), involving transparency (luminance) adjustment of chartgrid lines. Our second goal was to conduct additional ex-periments that demonstrate the use of Mechanical Turk forgenerating new insights. We studied rectangular area judg-ments (Exp. 1B), following the methodology of Cleveland &McGill to enable comparison, and then investigated optimalchart heights and gridline spacing (Exp. 3). Our third goalwas to analyze data from across our experiments to character-ize the use of Mechanical Turk as an experimental platform.
In the following four sections, we describe our experimentsand focus on details specific to visualization. Results of amore general nature are visited in our performance and costanalysis; for example, we delay discussion of response timeresults. Our experiments were initially launched with a lim-ited number of assignments (typically 3) to serve as a pilot.Upon completion of the trial assignments and verification ofthe results, the number of assignments was increased.
EXPERIMENT 1A: PROPORTIONAL JUDGMENTWe first replicated Cleveland & McGill’s seminal study [7]on Mechanical Turk. Their study was among the first to rankvisual variables empirically by their effectiveness for con-veying quantitative values. It also has influenced the designof automated presentation techniques [21, 22] and been suc-cessfully extended by others (e.g., [36]). As such, it is a nat-ural experiment to replicate to assess crowdsourcing.
MethodSeven judgment types, each corresponding to a visual en-coding (such as position or angle) were tested. The first fivecorrespond to Cleveland & McGill’s original position-lengthexperiment; types 1 through 3 use position encoding along acommon scale (Figure 1), while 4 and 5 use length encoding.Type 6 uses angle (as a pie chart) and type 7 uses circulararea (as a bubble chart, see Figure 2).
Ten charts were constructed at a resolution of 380⇥380 pix-els, for a total of 70 trials (HITs). We mimicked the number,values and aesthetics of the original charts as closely as pos-sible. For each chart, N=50 subjects were instructed first toidentify the smaller of two marked values, and then “makea quick visual judgment” to estimate what percentage thesmaller was of the larger. The first question served broadly toverify responses; only 14 out of 3,481 were incorrect (0.4%).Subjects were paid $0.05 per judgment.
To participate in the experiment, subjects first had to com-plete a qualification test consisting of two labeled examplecharts and three test charts. The test questions had the sameformat as the experiment trials, but with multiple choicerather than free text responses; only one choice was cor-rect, while the others were grossly wrong. The qualificationthus did not filter inaccurate subjects—which would bias theresponses—but ensured that subjects understood the instruc-tions. A pilot run of the experiment omitted this qualificationand over 10% of the responses were unusable. We discussthis observation in more detail later in the paper.
0
100
A B0
100
A B0
100
A B
Figure 1: Stimuli for judgment tasks T1, T2 & T3. Sub-jects estimated percent differences between elements.
A
B
B
A
A B
Figure 2: Area judgment stimuli. Top left: Bubblechart (T7), Bottom left: Center-aligned rectangles (T8),Right: Treemap (T9).
In the original experiment, Cleveland & McGill gave eachsubject a packet with all fifty charts on individual sheets.Lengthy tasks are ill-suited to Mechanical Turk; they aremore susceptible to “gaming” since the reward is higher, andsubjects cannot save drafts, raising the possibility of lost datadue to session timeout or connectivity error. We instead as-signed each chart as an individual task. Since the vast ma-jority (95%) of subjects accepted all tasks in sequence, theexperiment adhered to the original within-subjects format.
ResultsTo analyze responses, we replicated Cleveland & McGill’sdata exploration, using their log absolute error measure ofaccuracy: log2(|judged percent - true percent| + 1
8 ). We firstcomputed the midmeans of log absolute errors1 for each chart(Figure 3). The new results are similar (though not identical)to the originals: the rough shape and ranking of judgmenttypes by accuracy (T1-5) are preserved, supporting the valid-ity of the crowdsourced study.
Next we computed the log absolute error means and 95%confidence intervals for each judgment type using bootstrap-ping (c.f., [7]). The ranking of types by accuracy is consistentbetween the two experiments (Figure 4). Types 1 and 2 arecloser in the crowdsourced study; this may be a result of asmaller display mitigating the effect of distance. Types 4 and5 are more accurate than in the original study, but positionencoding still significantly outperformed length encoding.
We also introduced two new judgment types to evaluate an-gle and circular area encodings. Cleveland & McGill con-ducted a separate position-angle experiment; however, theyused a different task format, making it difficult to compare
1The midmean–the mean of the middle two quartiles–is a robust measureless susceptible to outliers. A log scale is used to measure relative propor-tional error and the 1
8 term is included to handle zero-valued differences.
Test % difference in length between elements
�18D. Koop, DSC 201, Fall 2018
[Heer & Bostock, 2010]
esting set of perceptual tasks, we replicated Cleveland &McGill’s [7] classic study (Exp. 1A) of proportionality es-timates across spatial encodings (position, length, angle),and Stone & Bartram’s [30] alpha contrast experiment (Exp.2), involving transparency (luminance) adjustment of chartgrid lines. Our second goal was to conduct additional ex-periments that demonstrate the use of Mechanical Turk forgenerating new insights. We studied rectangular area judg-ments (Exp. 1B), following the methodology of Cleveland &McGill to enable comparison, and then investigated optimalchart heights and gridline spacing (Exp. 3). Our third goalwas to analyze data from across our experiments to character-ize the use of Mechanical Turk as an experimental platform.
In the following four sections, we describe our experimentsand focus on details specific to visualization. Results of amore general nature are visited in our performance and costanalysis; for example, we delay discussion of response timeresults. Our experiments were initially launched with a lim-ited number of assignments (typically 3) to serve as a pilot.Upon completion of the trial assignments and verification ofthe results, the number of assignments was increased.
EXPERIMENT 1A: PROPORTIONAL JUDGMENTWe first replicated Cleveland & McGill’s seminal study [7]on Mechanical Turk. Their study was among the first to rankvisual variables empirically by their effectiveness for con-veying quantitative values. It also has influenced the designof automated presentation techniques [21, 22] and been suc-cessfully extended by others (e.g., [36]). As such, it is a nat-ural experiment to replicate to assess crowdsourcing.
MethodSeven judgment types, each corresponding to a visual en-coding (such as position or angle) were tested. The first fivecorrespond to Cleveland & McGill’s original position-lengthexperiment; types 1 through 3 use position encoding along acommon scale (Figure 1), while 4 and 5 use length encoding.Type 6 uses angle (as a pie chart) and type 7 uses circulararea (as a bubble chart, see Figure 2).
Ten charts were constructed at a resolution of 380⇥380 pix-els, for a total of 70 trials (HITs). We mimicked the number,values and aesthetics of the original charts as closely as pos-sible. For each chart, N=50 subjects were instructed first toidentify the smaller of two marked values, and then “makea quick visual judgment” to estimate what percentage thesmaller was of the larger. The first question served broadly toverify responses; only 14 out of 3,481 were incorrect (0.4%).Subjects were paid $0.05 per judgment.
To participate in the experiment, subjects first had to com-plete a qualification test consisting of two labeled examplecharts and three test charts. The test questions had the sameformat as the experiment trials, but with multiple choicerather than free text responses; only one choice was cor-rect, while the others were grossly wrong. The qualificationthus did not filter inaccurate subjects—which would bias theresponses—but ensured that subjects understood the instruc-tions. A pilot run of the experiment omitted this qualificationand over 10% of the responses were unusable. We discussthis observation in more detail later in the paper.
0
100
A B0
100
A B0
100
A B
Figure 1: Stimuli for judgment tasks T1, T2 & T3. Sub-jects estimated percent differences between elements.
A
B
B
A
A B
Figure 2: Area judgment stimuli. Top left: Bubblechart (T7), Bottom left: Center-aligned rectangles (T8),Right: Treemap (T9).
In the original experiment, Cleveland & McGill gave eachsubject a packet with all fifty charts on individual sheets.Lengthy tasks are ill-suited to Mechanical Turk; they aremore susceptible to “gaming” since the reward is higher, andsubjects cannot save drafts, raising the possibility of lost datadue to session timeout or connectivity error. We instead as-signed each chart as an individual task. Since the vast ma-jority (95%) of subjects accepted all tasks in sequence, theexperiment adhered to the original within-subjects format.
ResultsTo analyze responses, we replicated Cleveland & McGill’sdata exploration, using their log absolute error measure ofaccuracy: log2(|judged percent - true percent| + 1
8 ). We firstcomputed the midmeans of log absolute errors1 for each chart(Figure 3). The new results are similar (though not identical)to the originals: the rough shape and ranking of judgmenttypes by accuracy (T1-5) are preserved, supporting the valid-ity of the crowdsourced study.
Next we computed the log absolute error means and 95%confidence intervals for each judgment type using bootstrap-ping (c.f., [7]). The ranking of types by accuracy is consistentbetween the two experiments (Figure 4). Types 1 and 2 arecloser in the crowdsourced study; this may be a result of asmaller display mitigating the effect of distance. Types 4 and5 are more accurate than in the original study, but positionencoding still significantly outperformed length encoding.
We also introduced two new judgment types to evaluate an-gle and circular area encodings. Cleveland & McGill con-ducted a separate position-angle experiment; however, theyused a different task format, making it difficult to compare
1The midmean–the mean of the middle two quartiles–is a robust measureless susceptible to outliers. A log scale is used to measure relative propor-tional error and the 1
8 term is included to handle zero-valued differences.
Test % difference in length between elements
�19D. Koop, DSC 201, Fall 2018
[Heer & Bostock, 2010]
Answer: Left is ~5.6x longer than Right
esting set of perceptual tasks, we replicated Cleveland &McGill’s [7] classic study (Exp. 1A) of proportionality es-timates across spatial encodings (position, length, angle),and Stone & Bartram’s [30] alpha contrast experiment (Exp.2), involving transparency (luminance) adjustment of chartgrid lines. Our second goal was to conduct additional ex-periments that demonstrate the use of Mechanical Turk forgenerating new insights. We studied rectangular area judg-ments (Exp. 1B), following the methodology of Cleveland &McGill to enable comparison, and then investigated optimalchart heights and gridline spacing (Exp. 3). Our third goalwas to analyze data from across our experiments to character-ize the use of Mechanical Turk as an experimental platform.
In the following four sections, we describe our experimentsand focus on details specific to visualization. Results of amore general nature are visited in our performance and costanalysis; for example, we delay discussion of response timeresults. Our experiments were initially launched with a lim-ited number of assignments (typically 3) to serve as a pilot.Upon completion of the trial assignments and verification ofthe results, the number of assignments was increased.
EXPERIMENT 1A: PROPORTIONAL JUDGMENTWe first replicated Cleveland & McGill’s seminal study [7]on Mechanical Turk. Their study was among the first to rankvisual variables empirically by their effectiveness for con-veying quantitative values. It also has influenced the designof automated presentation techniques [21, 22] and been suc-cessfully extended by others (e.g., [36]). As such, it is a nat-ural experiment to replicate to assess crowdsourcing.
MethodSeven judgment types, each corresponding to a visual en-coding (such as position or angle) were tested. The first fivecorrespond to Cleveland & McGill’s original position-lengthexperiment; types 1 through 3 use position encoding along acommon scale (Figure 1), while 4 and 5 use length encoding.Type 6 uses angle (as a pie chart) and type 7 uses circulararea (as a bubble chart, see Figure 2).
Ten charts were constructed at a resolution of 380⇥380 pix-els, for a total of 70 trials (HITs). We mimicked the number,values and aesthetics of the original charts as closely as pos-sible. For each chart, N=50 subjects were instructed first toidentify the smaller of two marked values, and then “makea quick visual judgment” to estimate what percentage thesmaller was of the larger. The first question served broadly toverify responses; only 14 out of 3,481 were incorrect (0.4%).Subjects were paid $0.05 per judgment.
To participate in the experiment, subjects first had to com-plete a qualification test consisting of two labeled examplecharts and three test charts. The test questions had the sameformat as the experiment trials, but with multiple choicerather than free text responses; only one choice was cor-rect, while the others were grossly wrong. The qualificationthus did not filter inaccurate subjects—which would bias theresponses—but ensured that subjects understood the instruc-tions. A pilot run of the experiment omitted this qualificationand over 10% of the responses were unusable. We discussthis observation in more detail later in the paper.
0
100
A B0
100
A B0
100
A B
Figure 1: Stimuli for judgment tasks T1, T2 & T3. Sub-jects estimated percent differences between elements.
A
B
B
A
A B
Figure 2: Area judgment stimuli. Top left: Bubblechart (T7), Bottom left: Center-aligned rectangles (T8),Right: Treemap (T9).
In the original experiment, Cleveland & McGill gave eachsubject a packet with all fifty charts on individual sheets.Lengthy tasks are ill-suited to Mechanical Turk; they aremore susceptible to “gaming” since the reward is higher, andsubjects cannot save drafts, raising the possibility of lost datadue to session timeout or connectivity error. We instead as-signed each chart as an individual task. Since the vast ma-jority (95%) of subjects accepted all tasks in sequence, theexperiment adhered to the original within-subjects format.
ResultsTo analyze responses, we replicated Cleveland & McGill’sdata exploration, using their log absolute error measure ofaccuracy: log2(|judged percent - true percent| + 1
8 ). We firstcomputed the midmeans of log absolute errors1 for each chart(Figure 3). The new results are similar (though not identical)to the originals: the rough shape and ranking of judgmenttypes by accuracy (T1-5) are preserved, supporting the valid-ity of the crowdsourced study.
Next we computed the log absolute error means and 95%confidence intervals for each judgment type using bootstrap-ping (c.f., [7]). The ranking of types by accuracy is consistentbetween the two experiments (Figure 4). Types 1 and 2 arecloser in the crowdsourced study; this may be a result of asmaller display mitigating the effect of distance. Types 4 and5 are more accurate than in the original study, but positionencoding still significantly outperformed length encoding.
We also introduced two new judgment types to evaluate an-gle and circular area encodings. Cleveland & McGill con-ducted a separate position-angle experiment; however, theyused a different task format, making it difficult to compare
1The midmean–the mean of the middle two quartiles–is a robust measureless susceptible to outliers. A log scale is used to measure relative propor-tional error and the 1
8 term is included to handle zero-valued differences.
Test % difference in length between elements
�20D. Koop, DSC 201, Fall 2018
[Heer & Bostock, 2010]
esting set of perceptual tasks, we replicated Cleveland &McGill’s [7] classic study (Exp. 1A) of proportionality es-timates across spatial encodings (position, length, angle),and Stone & Bartram’s [30] alpha contrast experiment (Exp.2), involving transparency (luminance) adjustment of chartgrid lines. Our second goal was to conduct additional ex-periments that demonstrate the use of Mechanical Turk forgenerating new insights. We studied rectangular area judg-ments (Exp. 1B), following the methodology of Cleveland &McGill to enable comparison, and then investigated optimalchart heights and gridline spacing (Exp. 3). Our third goalwas to analyze data from across our experiments to character-ize the use of Mechanical Turk as an experimental platform.
In the following four sections, we describe our experimentsand focus on details specific to visualization. Results of amore general nature are visited in our performance and costanalysis; for example, we delay discussion of response timeresults. Our experiments were initially launched with a lim-ited number of assignments (typically 3) to serve as a pilot.Upon completion of the trial assignments and verification ofthe results, the number of assignments was increased.
EXPERIMENT 1A: PROPORTIONAL JUDGMENTWe first replicated Cleveland & McGill’s seminal study [7]on Mechanical Turk. Their study was among the first to rankvisual variables empirically by their effectiveness for con-veying quantitative values. It also has influenced the designof automated presentation techniques [21, 22] and been suc-cessfully extended by others (e.g., [36]). As such, it is a nat-ural experiment to replicate to assess crowdsourcing.
MethodSeven judgment types, each corresponding to a visual en-coding (such as position or angle) were tested. The first fivecorrespond to Cleveland & McGill’s original position-lengthexperiment; types 1 through 3 use position encoding along acommon scale (Figure 1), while 4 and 5 use length encoding.Type 6 uses angle (as a pie chart) and type 7 uses circulararea (as a bubble chart, see Figure 2).
Ten charts were constructed at a resolution of 380⇥380 pix-els, for a total of 70 trials (HITs). We mimicked the number,values and aesthetics of the original charts as closely as pos-sible. For each chart, N=50 subjects were instructed first toidentify the smaller of two marked values, and then “makea quick visual judgment” to estimate what percentage thesmaller was of the larger. The first question served broadly toverify responses; only 14 out of 3,481 were incorrect (0.4%).Subjects were paid $0.05 per judgment.
To participate in the experiment, subjects first had to com-plete a qualification test consisting of two labeled examplecharts and three test charts. The test questions had the sameformat as the experiment trials, but with multiple choicerather than free text responses; only one choice was cor-rect, while the others were grossly wrong. The qualificationthus did not filter inaccurate subjects—which would bias theresponses—but ensured that subjects understood the instruc-tions. A pilot run of the experiment omitted this qualificationand over 10% of the responses were unusable. We discussthis observation in more detail later in the paper.
0
100
A B0
100
A B0
100
A B
Figure 1: Stimuli for judgment tasks T1, T2 & T3. Sub-jects estimated percent differences between elements.
A
B
B
A
A B
Figure 2: Area judgment stimuli. Top left: Bubblechart (T7), Bottom left: Center-aligned rectangles (T8),Right: Treemap (T9).
In the original experiment, Cleveland & McGill gave eachsubject a packet with all fifty charts on individual sheets.Lengthy tasks are ill-suited to Mechanical Turk; they aremore susceptible to “gaming” since the reward is higher, andsubjects cannot save drafts, raising the possibility of lost datadue to session timeout or connectivity error. We instead as-signed each chart as an individual task. Since the vast ma-jority (95%) of subjects accepted all tasks in sequence, theexperiment adhered to the original within-subjects format.
ResultsTo analyze responses, we replicated Cleveland & McGill’sdata exploration, using their log absolute error measure ofaccuracy: log2(|judged percent - true percent| + 1
8 ). We firstcomputed the midmeans of log absolute errors1 for each chart(Figure 3). The new results are similar (though not identical)to the originals: the rough shape and ranking of judgmenttypes by accuracy (T1-5) are preserved, supporting the valid-ity of the crowdsourced study.
Next we computed the log absolute error means and 95%confidence intervals for each judgment type using bootstrap-ping (c.f., [7]). The ranking of types by accuracy is consistentbetween the two experiments (Figure 4). Types 1 and 2 arecloser in the crowdsourced study; this may be a result of asmaller display mitigating the effect of distance. Types 4 and5 are more accurate than in the original study, but positionencoding still significantly outperformed length encoding.
We also introduced two new judgment types to evaluate an-gle and circular area encodings. Cleveland & McGill con-ducted a separate position-angle experiment; however, theyused a different task format, making it difficult to compare
1The midmean–the mean of the middle two quartiles–is a robust measureless susceptible to outliers. A log scale is used to measure relative propor-tional error and the 1
8 term is included to handle zero-valued differences.
Test % difference in length between elements
�21D. Koop, DSC 201, Fall 2018
[Heer & Bostock, 2010]
esting set of perceptual tasks, we replicated Cleveland &McGill’s [7] classic study (Exp. 1A) of proportionality es-timates across spatial encodings (position, length, angle),and Stone & Bartram’s [30] alpha contrast experiment (Exp.2), involving transparency (luminance) adjustment of chartgrid lines. Our second goal was to conduct additional ex-periments that demonstrate the use of Mechanical Turk forgenerating new insights. We studied rectangular area judg-ments (Exp. 1B), following the methodology of Cleveland &McGill to enable comparison, and then investigated optimalchart heights and gridline spacing (Exp. 3). Our third goalwas to analyze data from across our experiments to character-ize the use of Mechanical Turk as an experimental platform.
In the following four sections, we describe our experimentsand focus on details specific to visualization. Results of amore general nature are visited in our performance and costanalysis; for example, we delay discussion of response timeresults. Our experiments were initially launched with a lim-ited number of assignments (typically 3) to serve as a pilot.Upon completion of the trial assignments and verification ofthe results, the number of assignments was increased.
EXPERIMENT 1A: PROPORTIONAL JUDGMENTWe first replicated Cleveland & McGill’s seminal study [7]on Mechanical Turk. Their study was among the first to rankvisual variables empirically by their effectiveness for con-veying quantitative values. It also has influenced the designof automated presentation techniques [21, 22] and been suc-cessfully extended by others (e.g., [36]). As such, it is a nat-ural experiment to replicate to assess crowdsourcing.
MethodSeven judgment types, each corresponding to a visual en-coding (such as position or angle) were tested. The first fivecorrespond to Cleveland & McGill’s original position-lengthexperiment; types 1 through 3 use position encoding along acommon scale (Figure 1), while 4 and 5 use length encoding.Type 6 uses angle (as a pie chart) and type 7 uses circulararea (as a bubble chart, see Figure 2).
Ten charts were constructed at a resolution of 380⇥380 pix-els, for a total of 70 trials (HITs). We mimicked the number,values and aesthetics of the original charts as closely as pos-sible. For each chart, N=50 subjects were instructed first toidentify the smaller of two marked values, and then “makea quick visual judgment” to estimate what percentage thesmaller was of the larger. The first question served broadly toverify responses; only 14 out of 3,481 were incorrect (0.4%).Subjects were paid $0.05 per judgment.
To participate in the experiment, subjects first had to com-plete a qualification test consisting of two labeled examplecharts and three test charts. The test questions had the sameformat as the experiment trials, but with multiple choicerather than free text responses; only one choice was cor-rect, while the others were grossly wrong. The qualificationthus did not filter inaccurate subjects—which would bias theresponses—but ensured that subjects understood the instruc-tions. A pilot run of the experiment omitted this qualificationand over 10% of the responses were unusable. We discussthis observation in more detail later in the paper.
0
100
A B0
100
A B0
100
A B
Figure 1: Stimuli for judgment tasks T1, T2 & T3. Sub-jects estimated percent differences between elements.
A
B
B
A
A B
Figure 2: Area judgment stimuli. Top left: Bubblechart (T7), Bottom left: Center-aligned rectangles (T8),Right: Treemap (T9).
In the original experiment, Cleveland & McGill gave eachsubject a packet with all fifty charts on individual sheets.Lengthy tasks are ill-suited to Mechanical Turk; they aremore susceptible to “gaming” since the reward is higher, andsubjects cannot save drafts, raising the possibility of lost datadue to session timeout or connectivity error. We instead as-signed each chart as an individual task. Since the vast ma-jority (95%) of subjects accepted all tasks in sequence, theexperiment adhered to the original within-subjects format.
ResultsTo analyze responses, we replicated Cleveland & McGill’sdata exploration, using their log absolute error measure ofaccuracy: log2(|judged percent - true percent| + 1
8 ). We firstcomputed the midmeans of log absolute errors1 for each chart(Figure 3). The new results are similar (though not identical)to the originals: the rough shape and ranking of judgmenttypes by accuracy (T1-5) are preserved, supporting the valid-ity of the crowdsourced study.
Next we computed the log absolute error means and 95%confidence intervals for each judgment type using bootstrap-ping (c.f., [7]). The ranking of types by accuracy is consistentbetween the two experiments (Figure 4). Types 1 and 2 arecloser in the crowdsourced study; this may be a result of asmaller display mitigating the effect of distance. Types 4 and5 are more accurate than in the original study, but positionencoding still significantly outperformed length encoding.
We also introduced two new judgment types to evaluate an-gle and circular area encodings. Cleveland & McGill con-ducted a separate position-angle experiment; however, theyused a different task format, making it difficult to compare
1The midmean–the mean of the middle two quartiles–is a robust measureless susceptible to outliers. A log scale is used to measure relative propor-tional error and the 1
8 term is included to handle zero-valued differences.
Test % difference in length between elements
�22D. Koop, DSC 201, Fall 2018
[Modified from Heer & Bostock, 2010]
esting set of perceptual tasks, we replicated Cleveland &McGill’s [7] classic study (Exp. 1A) of proportionality es-timates across spatial encodings (position, length, angle),and Stone & Bartram’s [30] alpha contrast experiment (Exp.2), involving transparency (luminance) adjustment of chartgrid lines. Our second goal was to conduct additional ex-periments that demonstrate the use of Mechanical Turk forgenerating new insights. We studied rectangular area judg-ments (Exp. 1B), following the methodology of Cleveland &McGill to enable comparison, and then investigated optimalchart heights and gridline spacing (Exp. 3). Our third goalwas to analyze data from across our experiments to character-ize the use of Mechanical Turk as an experimental platform.
In the following four sections, we describe our experimentsand focus on details specific to visualization. Results of amore general nature are visited in our performance and costanalysis; for example, we delay discussion of response timeresults. Our experiments were initially launched with a lim-ited number of assignments (typically 3) to serve as a pilot.Upon completion of the trial assignments and verification ofthe results, the number of assignments was increased.
EXPERIMENT 1A: PROPORTIONAL JUDGMENTWe first replicated Cleveland & McGill’s seminal study [7]on Mechanical Turk. Their study was among the first to rankvisual variables empirically by their effectiveness for con-veying quantitative values. It also has influenced the designof automated presentation techniques [21, 22] and been suc-cessfully extended by others (e.g., [36]). As such, it is a nat-ural experiment to replicate to assess crowdsourcing.
MethodSeven judgment types, each corresponding to a visual en-coding (such as position or angle) were tested. The first fivecorrespond to Cleveland & McGill’s original position-lengthexperiment; types 1 through 3 use position encoding along acommon scale (Figure 1), while 4 and 5 use length encoding.Type 6 uses angle (as a pie chart) and type 7 uses circulararea (as a bubble chart, see Figure 2).
Ten charts were constructed at a resolution of 380⇥380 pix-els, for a total of 70 trials (HITs). We mimicked the number,values and aesthetics of the original charts as closely as pos-sible. For each chart, N=50 subjects were instructed first toidentify the smaller of two marked values, and then “makea quick visual judgment” to estimate what percentage thesmaller was of the larger. The first question served broadly toverify responses; only 14 out of 3,481 were incorrect (0.4%).Subjects were paid $0.05 per judgment.
To participate in the experiment, subjects first had to com-plete a qualification test consisting of two labeled examplecharts and three test charts. The test questions had the sameformat as the experiment trials, but with multiple choicerather than free text responses; only one choice was cor-rect, while the others were grossly wrong. The qualificationthus did not filter inaccurate subjects—which would bias theresponses—but ensured that subjects understood the instruc-tions. A pilot run of the experiment omitted this qualificationand over 10% of the responses were unusable. We discussthis observation in more detail later in the paper.
0
100
A B0
100
A B0
100
A B
Figure 1: Stimuli for judgment tasks T1, T2 & T3. Sub-jects estimated percent differences between elements.
A
B
B
A
A B
Figure 2: Area judgment stimuli. Top left: Bubblechart (T7), Bottom left: Center-aligned rectangles (T8),Right: Treemap (T9).
In the original experiment, Cleveland & McGill gave eachsubject a packet with all fifty charts on individual sheets.Lengthy tasks are ill-suited to Mechanical Turk; they aremore susceptible to “gaming” since the reward is higher, andsubjects cannot save drafts, raising the possibility of lost datadue to session timeout or connectivity error. We instead as-signed each chart as an individual task. Since the vast ma-jority (95%) of subjects accepted all tasks in sequence, theexperiment adhered to the original within-subjects format.
ResultsTo analyze responses, we replicated Cleveland & McGill’sdata exploration, using their log absolute error measure ofaccuracy: log2(|judged percent - true percent| + 1
8 ). We firstcomputed the midmeans of log absolute errors1 for each chart(Figure 3). The new results are similar (though not identical)to the originals: the rough shape and ranking of judgmenttypes by accuracy (T1-5) are preserved, supporting the valid-ity of the crowdsourced study.
Next we computed the log absolute error means and 95%confidence intervals for each judgment type using bootstrap-ping (c.f., [7]). The ranking of types by accuracy is consistentbetween the two experiments (Figure 4). Types 1 and 2 arecloser in the crowdsourced study; this may be a result of asmaller display mitigating the effect of distance. Types 4 and5 are more accurate than in the original study, but positionencoding still significantly outperformed length encoding.
We also introduced two new judgment types to evaluate an-gle and circular area encodings. Cleveland & McGill con-ducted a separate position-angle experiment; however, theyused a different task format, making it difficult to compare
1The midmean–the mean of the middle two quartiles–is a robust measureless susceptible to outliers. A log scale is used to measure relative propor-tional error and the 1
8 term is included to handle zero-valued differences.
Test % difference in length between elements
�23D. Koop, DSC 201, Fall 2018
[Modified from Heer & Bostock, 2010]
Answer: Right is 4x larger than Left
esting set of perceptual tasks, we replicated Cleveland &McGill’s [7] classic study (Exp. 1A) of proportionality es-timates across spatial encodings (position, length, angle),and Stone & Bartram’s [30] alpha contrast experiment (Exp.2), involving transparency (luminance) adjustment of chartgrid lines. Our second goal was to conduct additional ex-periments that demonstrate the use of Mechanical Turk forgenerating new insights. We studied rectangular area judg-ments (Exp. 1B), following the methodology of Cleveland &McGill to enable comparison, and then investigated optimalchart heights and gridline spacing (Exp. 3). Our third goalwas to analyze data from across our experiments to character-ize the use of Mechanical Turk as an experimental platform.
In the following four sections, we describe our experimentsand focus on details specific to visualization. Results of amore general nature are visited in our performance and costanalysis; for example, we delay discussion of response timeresults. Our experiments were initially launched with a lim-ited number of assignments (typically 3) to serve as a pilot.Upon completion of the trial assignments and verification ofthe results, the number of assignments was increased.
EXPERIMENT 1A: PROPORTIONAL JUDGMENTWe first replicated Cleveland & McGill’s seminal study [7]on Mechanical Turk. Their study was among the first to rankvisual variables empirically by their effectiveness for con-veying quantitative values. It also has influenced the designof automated presentation techniques [21, 22] and been suc-cessfully extended by others (e.g., [36]). As such, it is a nat-ural experiment to replicate to assess crowdsourcing.
MethodSeven judgment types, each corresponding to a visual en-coding (such as position or angle) were tested. The first fivecorrespond to Cleveland & McGill’s original position-lengthexperiment; types 1 through 3 use position encoding along acommon scale (Figure 1), while 4 and 5 use length encoding.Type 6 uses angle (as a pie chart) and type 7 uses circulararea (as a bubble chart, see Figure 2).
Ten charts were constructed at a resolution of 380⇥380 pix-els, for a total of 70 trials (HITs). We mimicked the number,values and aesthetics of the original charts as closely as pos-sible. For each chart, N=50 subjects were instructed first toidentify the smaller of two marked values, and then “makea quick visual judgment” to estimate what percentage thesmaller was of the larger. The first question served broadly toverify responses; only 14 out of 3,481 were incorrect (0.4%).Subjects were paid $0.05 per judgment.
To participate in the experiment, subjects first had to com-plete a qualification test consisting of two labeled examplecharts and three test charts. The test questions had the sameformat as the experiment trials, but with multiple choicerather than free text responses; only one choice was cor-rect, while the others were grossly wrong. The qualificationthus did not filter inaccurate subjects—which would bias theresponses—but ensured that subjects understood the instruc-tions. A pilot run of the experiment omitted this qualificationand over 10% of the responses were unusable. We discussthis observation in more detail later in the paper.
0
100
A B0
100
A B0
100
A B
Figure 1: Stimuli for judgment tasks T1, T2 & T3. Sub-jects estimated percent differences between elements.
A
B
B
A
A B
Figure 2: Area judgment stimuli. Top left: Bubblechart (T7), Bottom left: Center-aligned rectangles (T8),Right: Treemap (T9).
In the original experiment, Cleveland & McGill gave eachsubject a packet with all fifty charts on individual sheets.Lengthy tasks are ill-suited to Mechanical Turk; they aremore susceptible to “gaming” since the reward is higher, andsubjects cannot save drafts, raising the possibility of lost datadue to session timeout or connectivity error. We instead as-signed each chart as an individual task. Since the vast ma-jority (95%) of subjects accepted all tasks in sequence, theexperiment adhered to the original within-subjects format.
ResultsTo analyze responses, we replicated Cleveland & McGill’sdata exploration, using their log absolute error measure ofaccuracy: log2(|judged percent - true percent| + 1
8 ). We firstcomputed the midmeans of log absolute errors1 for each chart(Figure 3). The new results are similar (though not identical)to the originals: the rough shape and ranking of judgmenttypes by accuracy (T1-5) are preserved, supporting the valid-ity of the crowdsourced study.
Next we computed the log absolute error means and 95%confidence intervals for each judgment type using bootstrap-ping (c.f., [7]). The ranking of types by accuracy is consistentbetween the two experiments (Figure 4). Types 1 and 2 arecloser in the crowdsourced study; this may be a result of asmaller display mitigating the effect of distance. Types 4 and5 are more accurate than in the original study, but positionencoding still significantly outperformed length encoding.
We also introduced two new judgment types to evaluate an-gle and circular area encodings. Cleveland & McGill con-ducted a separate position-angle experiment; however, theyused a different task format, making it difficult to compare
1The midmean–the mean of the middle two quartiles–is a robust measureless susceptible to outliers. A log scale is used to measure relative propor-tional error and the 1
8 term is included to handle zero-valued differences.
Test % difference in area between elements
�24D. Koop, DSC 201, Fall 2018
[Heer & Bostock, 2010]
esting set of perceptual tasks, we replicated Cleveland &McGill’s [7] classic study (Exp. 1A) of proportionality es-timates across spatial encodings (position, length, angle),and Stone & Bartram’s [30] alpha contrast experiment (Exp.2), involving transparency (luminance) adjustment of chartgrid lines. Our second goal was to conduct additional ex-periments that demonstrate the use of Mechanical Turk forgenerating new insights. We studied rectangular area judg-ments (Exp. 1B), following the methodology of Cleveland &McGill to enable comparison, and then investigated optimalchart heights and gridline spacing (Exp. 3). Our third goalwas to analyze data from across our experiments to character-ize the use of Mechanical Turk as an experimental platform.
In the following four sections, we describe our experimentsand focus on details specific to visualization. Results of amore general nature are visited in our performance and costanalysis; for example, we delay discussion of response timeresults. Our experiments were initially launched with a lim-ited number of assignments (typically 3) to serve as a pilot.Upon completion of the trial assignments and verification ofthe results, the number of assignments was increased.
EXPERIMENT 1A: PROPORTIONAL JUDGMENTWe first replicated Cleveland & McGill’s seminal study [7]on Mechanical Turk. Their study was among the first to rankvisual variables empirically by their effectiveness for con-veying quantitative values. It also has influenced the designof automated presentation techniques [21, 22] and been suc-cessfully extended by others (e.g., [36]). As such, it is a nat-ural experiment to replicate to assess crowdsourcing.
MethodSeven judgment types, each corresponding to a visual en-coding (such as position or angle) were tested. The first fivecorrespond to Cleveland & McGill’s original position-lengthexperiment; types 1 through 3 use position encoding along acommon scale (Figure 1), while 4 and 5 use length encoding.Type 6 uses angle (as a pie chart) and type 7 uses circulararea (as a bubble chart, see Figure 2).
Ten charts were constructed at a resolution of 380⇥380 pix-els, for a total of 70 trials (HITs). We mimicked the number,values and aesthetics of the original charts as closely as pos-sible. For each chart, N=50 subjects were instructed first toidentify the smaller of two marked values, and then “makea quick visual judgment” to estimate what percentage thesmaller was of the larger. The first question served broadly toverify responses; only 14 out of 3,481 were incorrect (0.4%).Subjects were paid $0.05 per judgment.
To participate in the experiment, subjects first had to com-plete a qualification test consisting of two labeled examplecharts and three test charts. The test questions had the sameformat as the experiment trials, but with multiple choicerather than free text responses; only one choice was cor-rect, while the others were grossly wrong. The qualificationthus did not filter inaccurate subjects—which would bias theresponses—but ensured that subjects understood the instruc-tions. A pilot run of the experiment omitted this qualificationand over 10% of the responses were unusable. We discussthis observation in more detail later in the paper.
0
100
A B0
100
A B0
100
A B
Figure 1: Stimuli for judgment tasks T1, T2 & T3. Sub-jects estimated percent differences between elements.
A
B
B
A
A B
Figure 2: Area judgment stimuli. Top left: Bubblechart (T7), Bottom left: Center-aligned rectangles (T8),Right: Treemap (T9).
In the original experiment, Cleveland & McGill gave eachsubject a packet with all fifty charts on individual sheets.Lengthy tasks are ill-suited to Mechanical Turk; they aremore susceptible to “gaming” since the reward is higher, andsubjects cannot save drafts, raising the possibility of lost datadue to session timeout or connectivity error. We instead as-signed each chart as an individual task. Since the vast ma-jority (95%) of subjects accepted all tasks in sequence, theexperiment adhered to the original within-subjects format.
ResultsTo analyze responses, we replicated Cleveland & McGill’sdata exploration, using their log absolute error measure ofaccuracy: log2(|judged percent - true percent| + 1
8 ). We firstcomputed the midmeans of log absolute errors1 for each chart(Figure 3). The new results are similar (though not identical)to the originals: the rough shape and ranking of judgmenttypes by accuracy (T1-5) are preserved, supporting the valid-ity of the crowdsourced study.
Next we computed the log absolute error means and 95%confidence intervals for each judgment type using bootstrap-ping (c.f., [7]). The ranking of types by accuracy is consistentbetween the two experiments (Figure 4). Types 1 and 2 arecloser in the crowdsourced study; this may be a result of asmaller display mitigating the effect of distance. Types 4 and5 are more accurate than in the original study, but positionencoding still significantly outperformed length encoding.
We also introduced two new judgment types to evaluate an-gle and circular area encodings. Cleveland & McGill con-ducted a separate position-angle experiment; however, theyused a different task format, making it difficult to compare
1The midmean–the mean of the middle two quartiles–is a robust measureless susceptible to outliers. A log scale is used to measure relative propor-tional error and the 1
8 term is included to handle zero-valued differences.
Test % difference in area between elements
�25D. Koop, DSC 201, Fall 2018
[Heer & Bostock, 2010]
Answer: A is ~2.25x larger (in area) than B
esting set of perceptual tasks, we replicated Cleveland &McGill’s [7] classic study (Exp. 1A) of proportionality es-timates across spatial encodings (position, length, angle),and Stone & Bartram’s [30] alpha contrast experiment (Exp.2), involving transparency (luminance) adjustment of chartgrid lines. Our second goal was to conduct additional ex-periments that demonstrate the use of Mechanical Turk forgenerating new insights. We studied rectangular area judg-ments (Exp. 1B), following the methodology of Cleveland &McGill to enable comparison, and then investigated optimalchart heights and gridline spacing (Exp. 3). Our third goalwas to analyze data from across our experiments to character-ize the use of Mechanical Turk as an experimental platform.
In the following four sections, we describe our experimentsand focus on details specific to visualization. Results of amore general nature are visited in our performance and costanalysis; for example, we delay discussion of response timeresults. Our experiments were initially launched with a lim-ited number of assignments (typically 3) to serve as a pilot.Upon completion of the trial assignments and verification ofthe results, the number of assignments was increased.
EXPERIMENT 1A: PROPORTIONAL JUDGMENTWe first replicated Cleveland & McGill’s seminal study [7]on Mechanical Turk. Their study was among the first to rankvisual variables empirically by their effectiveness for con-veying quantitative values. It also has influenced the designof automated presentation techniques [21, 22] and been suc-cessfully extended by others (e.g., [36]). As such, it is a nat-ural experiment to replicate to assess crowdsourcing.
MethodSeven judgment types, each corresponding to a visual en-coding (such as position or angle) were tested. The first fivecorrespond to Cleveland & McGill’s original position-lengthexperiment; types 1 through 3 use position encoding along acommon scale (Figure 1), while 4 and 5 use length encoding.Type 6 uses angle (as a pie chart) and type 7 uses circulararea (as a bubble chart, see Figure 2).
Ten charts were constructed at a resolution of 380⇥380 pix-els, for a total of 70 trials (HITs). We mimicked the number,values and aesthetics of the original charts as closely as pos-sible. For each chart, N=50 subjects were instructed first toidentify the smaller of two marked values, and then “makea quick visual judgment” to estimate what percentage thesmaller was of the larger. The first question served broadly toverify responses; only 14 out of 3,481 were incorrect (0.4%).Subjects were paid $0.05 per judgment.
To participate in the experiment, subjects first had to com-plete a qualification test consisting of two labeled examplecharts and three test charts. The test questions had the sameformat as the experiment trials, but with multiple choicerather than free text responses; only one choice was cor-rect, while the others were grossly wrong. The qualificationthus did not filter inaccurate subjects—which would bias theresponses—but ensured that subjects understood the instruc-tions. A pilot run of the experiment omitted this qualificationand over 10% of the responses were unusable. We discussthis observation in more detail later in the paper.
0
100
A B0
100
A B0
100
A B
Figure 1: Stimuli for judgment tasks T1, T2 & T3. Sub-jects estimated percent differences between elements.
A
B
B
A
A B
Figure 2: Area judgment stimuli. Top left: Bubblechart (T7), Bottom left: Center-aligned rectangles (T8),Right: Treemap (T9).
In the original experiment, Cleveland & McGill gave eachsubject a packet with all fifty charts on individual sheets.Lengthy tasks are ill-suited to Mechanical Turk; they aremore susceptible to “gaming” since the reward is higher, andsubjects cannot save drafts, raising the possibility of lost datadue to session timeout or connectivity error. We instead as-signed each chart as an individual task. Since the vast ma-jority (95%) of subjects accepted all tasks in sequence, theexperiment adhered to the original within-subjects format.
ResultsTo analyze responses, we replicated Cleveland & McGill’sdata exploration, using their log absolute error measure ofaccuracy: log2(|judged percent - true percent| + 1
8 ). We firstcomputed the midmeans of log absolute errors1 for each chart(Figure 3). The new results are similar (though not identical)to the originals: the rough shape and ranking of judgmenttypes by accuracy (T1-5) are preserved, supporting the valid-ity of the crowdsourced study.
Next we computed the log absolute error means and 95%confidence intervals for each judgment type using bootstrap-ping (c.f., [7]). The ranking of types by accuracy is consistentbetween the two experiments (Figure 4). Types 1 and 2 arecloser in the crowdsourced study; this may be a result of asmaller display mitigating the effect of distance. Types 4 and5 are more accurate than in the original study, but positionencoding still significantly outperformed length encoding.
We also introduced two new judgment types to evaluate an-gle and circular area encodings. Cleveland & McGill con-ducted a separate position-angle experiment; however, theyused a different task format, making it difficult to compare
1The midmean–the mean of the middle two quartiles–is a robust measureless susceptible to outliers. A log scale is used to measure relative propor-tional error and the 1
8 term is included to handle zero-valued differences.
Test % difference in area between elements
�26D. Koop, DSC 201, Fall 2018
[Heer & Bostock, 2010]
esting set of perceptual tasks, we replicated Cleveland &McGill’s [7] classic study (Exp. 1A) of proportionality es-timates across spatial encodings (position, length, angle),and Stone & Bartram’s [30] alpha contrast experiment (Exp.2), involving transparency (luminance) adjustment of chartgrid lines. Our second goal was to conduct additional ex-periments that demonstrate the use of Mechanical Turk forgenerating new insights. We studied rectangular area judg-ments (Exp. 1B), following the methodology of Cleveland &McGill to enable comparison, and then investigated optimalchart heights and gridline spacing (Exp. 3). Our third goalwas to analyze data from across our experiments to character-ize the use of Mechanical Turk as an experimental platform.
In the following four sections, we describe our experimentsand focus on details specific to visualization. Results of amore general nature are visited in our performance and costanalysis; for example, we delay discussion of response timeresults. Our experiments were initially launched with a lim-ited number of assignments (typically 3) to serve as a pilot.Upon completion of the trial assignments and verification ofthe results, the number of assignments was increased.
EXPERIMENT 1A: PROPORTIONAL JUDGMENTWe first replicated Cleveland & McGill’s seminal study [7]on Mechanical Turk. Their study was among the first to rankvisual variables empirically by their effectiveness for con-veying quantitative values. It also has influenced the designof automated presentation techniques [21, 22] and been suc-cessfully extended by others (e.g., [36]). As such, it is a nat-ural experiment to replicate to assess crowdsourcing.
MethodSeven judgment types, each corresponding to a visual en-coding (such as position or angle) were tested. The first fivecorrespond to Cleveland & McGill’s original position-lengthexperiment; types 1 through 3 use position encoding along acommon scale (Figure 1), while 4 and 5 use length encoding.Type 6 uses angle (as a pie chart) and type 7 uses circulararea (as a bubble chart, see Figure 2).
Ten charts were constructed at a resolution of 380⇥380 pix-els, for a total of 70 trials (HITs). We mimicked the number,values and aesthetics of the original charts as closely as pos-sible. For each chart, N=50 subjects were instructed first toidentify the smaller of two marked values, and then “makea quick visual judgment” to estimate what percentage thesmaller was of the larger. The first question served broadly toverify responses; only 14 out of 3,481 were incorrect (0.4%).Subjects were paid $0.05 per judgment.
To participate in the experiment, subjects first had to com-plete a qualification test consisting of two labeled examplecharts and three test charts. The test questions had the sameformat as the experiment trials, but with multiple choicerather than free text responses; only one choice was cor-rect, while the others were grossly wrong. The qualificationthus did not filter inaccurate subjects—which would bias theresponses—but ensured that subjects understood the instruc-tions. A pilot run of the experiment omitted this qualificationand over 10% of the responses were unusable. We discussthis observation in more detail later in the paper.
0
100
A B0
100
A B0
100
A B
Figure 1: Stimuli for judgment tasks T1, T2 & T3. Sub-jects estimated percent differences between elements.
A
B
B
A
A B
Figure 2: Area judgment stimuli. Top left: Bubblechart (T7), Bottom left: Center-aligned rectangles (T8),Right: Treemap (T9).
In the original experiment, Cleveland & McGill gave eachsubject a packet with all fifty charts on individual sheets.Lengthy tasks are ill-suited to Mechanical Turk; they aremore susceptible to “gaming” since the reward is higher, andsubjects cannot save drafts, raising the possibility of lost datadue to session timeout or connectivity error. We instead as-signed each chart as an individual task. Since the vast ma-jority (95%) of subjects accepted all tasks in sequence, theexperiment adhered to the original within-subjects format.
ResultsTo analyze responses, we replicated Cleveland & McGill’sdata exploration, using their log absolute error measure ofaccuracy: log2(|judged percent - true percent| + 1
8 ). We firstcomputed the midmeans of log absolute errors1 for each chart(Figure 3). The new results are similar (though not identical)to the originals: the rough shape and ranking of judgmenttypes by accuracy (T1-5) are preserved, supporting the valid-ity of the crowdsourced study.
Next we computed the log absolute error means and 95%confidence intervals for each judgment type using bootstrap-ping (c.f., [7]). The ranking of types by accuracy is consistentbetween the two experiments (Figure 4). Types 1 and 2 arecloser in the crowdsourced study; this may be a result of asmaller display mitigating the effect of distance. Types 4 and5 are more accurate than in the original study, but positionencoding still significantly outperformed length encoding.
We also introduced two new judgment types to evaluate an-gle and circular area encodings. Cleveland & McGill con-ducted a separate position-angle experiment; however, theyused a different task format, making it difficult to compare
1The midmean–the mean of the middle two quartiles–is a robust measureless susceptible to outliers. A log scale is used to measure relative propor-tional error and the 1
8 term is included to handle zero-valued differences.
Test % difference in area between elements
�27D. Koop, DSC 201, Fall 2018
[Heer & Bostock, 2010]
Answer: B is ~6.1x larger (in area) than A
esting set of perceptual tasks, we replicated Cleveland &McGill’s [7] classic study (Exp. 1A) of proportionality es-timates across spatial encodings (position, length, angle),and Stone & Bartram’s [30] alpha contrast experiment (Exp.2), involving transparency (luminance) adjustment of chartgrid lines. Our second goal was to conduct additional ex-periments that demonstrate the use of Mechanical Turk forgenerating new insights. We studied rectangular area judg-ments (Exp. 1B), following the methodology of Cleveland &McGill to enable comparison, and then investigated optimalchart heights and gridline spacing (Exp. 3). Our third goalwas to analyze data from across our experiments to character-ize the use of Mechanical Turk as an experimental platform.
In the following four sections, we describe our experimentsand focus on details specific to visualization. Results of amore general nature are visited in our performance and costanalysis; for example, we delay discussion of response timeresults. Our experiments were initially launched with a lim-ited number of assignments (typically 3) to serve as a pilot.Upon completion of the trial assignments and verification ofthe results, the number of assignments was increased.
EXPERIMENT 1A: PROPORTIONAL JUDGMENTWe first replicated Cleveland & McGill’s seminal study [7]on Mechanical Turk. Their study was among the first to rankvisual variables empirically by their effectiveness for con-veying quantitative values. It also has influenced the designof automated presentation techniques [21, 22] and been suc-cessfully extended by others (e.g., [36]). As such, it is a nat-ural experiment to replicate to assess crowdsourcing.
MethodSeven judgment types, each corresponding to a visual en-coding (such as position or angle) were tested. The first fivecorrespond to Cleveland & McGill’s original position-lengthexperiment; types 1 through 3 use position encoding along acommon scale (Figure 1), while 4 and 5 use length encoding.Type 6 uses angle (as a pie chart) and type 7 uses circulararea (as a bubble chart, see Figure 2).
Ten charts were constructed at a resolution of 380⇥380 pix-els, for a total of 70 trials (HITs). We mimicked the number,values and aesthetics of the original charts as closely as pos-sible. For each chart, N=50 subjects were instructed first toidentify the smaller of two marked values, and then “makea quick visual judgment” to estimate what percentage thesmaller was of the larger. The first question served broadly toverify responses; only 14 out of 3,481 were incorrect (0.4%).Subjects were paid $0.05 per judgment.
To participate in the experiment, subjects first had to com-plete a qualification test consisting of two labeled examplecharts and three test charts. The test questions had the sameformat as the experiment trials, but with multiple choicerather than free text responses; only one choice was cor-rect, while the others were grossly wrong. The qualificationthus did not filter inaccurate subjects—which would bias theresponses—but ensured that subjects understood the instruc-tions. A pilot run of the experiment omitted this qualificationand over 10% of the responses were unusable. We discussthis observation in more detail later in the paper.
0
100
A B0
100
A B0
100
A B
Figure 1: Stimuli for judgment tasks T1, T2 & T3. Sub-jects estimated percent differences between elements.
A
B
B
A
A B
Figure 2: Area judgment stimuli. Top left: Bubblechart (T7), Bottom left: Center-aligned rectangles (T8),Right: Treemap (T9).
In the original experiment, Cleveland & McGill gave eachsubject a packet with all fifty charts on individual sheets.Lengthy tasks are ill-suited to Mechanical Turk; they aremore susceptible to “gaming” since the reward is higher, andsubjects cannot save drafts, raising the possibility of lost datadue to session timeout or connectivity error. We instead as-signed each chart as an individual task. Since the vast ma-jority (95%) of subjects accepted all tasks in sequence, theexperiment adhered to the original within-subjects format.
ResultsTo analyze responses, we replicated Cleveland & McGill’sdata exploration, using their log absolute error measure ofaccuracy: log2(|judged percent - true percent| + 1
8 ). We firstcomputed the midmeans of log absolute errors1 for each chart(Figure 3). The new results are similar (though not identical)to the originals: the rough shape and ranking of judgmenttypes by accuracy (T1-5) are preserved, supporting the valid-ity of the crowdsourced study.
Next we computed the log absolute error means and 95%confidence intervals for each judgment type using bootstrap-ping (c.f., [7]). The ranking of types by accuracy is consistentbetween the two experiments (Figure 4). Types 1 and 2 arecloser in the crowdsourced study; this may be a result of asmaller display mitigating the effect of distance. Types 4 and5 are more accurate than in the original study, but positionencoding still significantly outperformed length encoding.
We also introduced two new judgment types to evaluate an-gle and circular area encodings. Cleveland & McGill con-ducted a separate position-angle experiment; however, theyused a different task format, making it difficult to compare
1The midmean–the mean of the middle two quartiles–is a robust measureless susceptible to outliers. A log scale is used to measure relative propor-tional error and the 1
8 term is included to handle zero-valued differences.
Test % difference in area between elements
�28D. Koop, DSC 201, Fall 2018
[Heer & Bostock, 2010]
esting set of perceptual tasks, we replicated Cleveland &McGill’s [7] classic study (Exp. 1A) of proportionality es-timates across spatial encodings (position, length, angle),and Stone & Bartram’s [30] alpha contrast experiment (Exp.2), involving transparency (luminance) adjustment of chartgrid lines. Our second goal was to conduct additional ex-periments that demonstrate the use of Mechanical Turk forgenerating new insights. We studied rectangular area judg-ments (Exp. 1B), following the methodology of Cleveland &McGill to enable comparison, and then investigated optimalchart heights and gridline spacing (Exp. 3). Our third goalwas to analyze data from across our experiments to character-ize the use of Mechanical Turk as an experimental platform.
In the following four sections, we describe our experimentsand focus on details specific to visualization. Results of amore general nature are visited in our performance and costanalysis; for example, we delay discussion of response timeresults. Our experiments were initially launched with a lim-ited number of assignments (typically 3) to serve as a pilot.Upon completion of the trial assignments and verification ofthe results, the number of assignments was increased.
EXPERIMENT 1A: PROPORTIONAL JUDGMENTWe first replicated Cleveland & McGill’s seminal study [7]on Mechanical Turk. Their study was among the first to rankvisual variables empirically by their effectiveness for con-veying quantitative values. It also has influenced the designof automated presentation techniques [21, 22] and been suc-cessfully extended by others (e.g., [36]). As such, it is a nat-ural experiment to replicate to assess crowdsourcing.
MethodSeven judgment types, each corresponding to a visual en-coding (such as position or angle) were tested. The first fivecorrespond to Cleveland & McGill’s original position-lengthexperiment; types 1 through 3 use position encoding along acommon scale (Figure 1), while 4 and 5 use length encoding.Type 6 uses angle (as a pie chart) and type 7 uses circulararea (as a bubble chart, see Figure 2).
Ten charts were constructed at a resolution of 380⇥380 pix-els, for a total of 70 trials (HITs). We mimicked the number,values and aesthetics of the original charts as closely as pos-sible. For each chart, N=50 subjects were instructed first toidentify the smaller of two marked values, and then “makea quick visual judgment” to estimate what percentage thesmaller was of the larger. The first question served broadly toverify responses; only 14 out of 3,481 were incorrect (0.4%).Subjects were paid $0.05 per judgment.
To participate in the experiment, subjects first had to com-plete a qualification test consisting of two labeled examplecharts and three test charts. The test questions had the sameformat as the experiment trials, but with multiple choicerather than free text responses; only one choice was cor-rect, while the others were grossly wrong. The qualificationthus did not filter inaccurate subjects—which would bias theresponses—but ensured that subjects understood the instruc-tions. A pilot run of the experiment omitted this qualificationand over 10% of the responses were unusable. We discussthis observation in more detail later in the paper.
0
100
A B0
100
A B0
100
A B
Figure 1: Stimuli for judgment tasks T1, T2 & T3. Sub-jects estimated percent differences between elements.
A
B
B
A
A B
Figure 2: Area judgment stimuli. Top left: Bubblechart (T7), Bottom left: Center-aligned rectangles (T8),Right: Treemap (T9).
In the original experiment, Cleveland & McGill gave eachsubject a packet with all fifty charts on individual sheets.Lengthy tasks are ill-suited to Mechanical Turk; they aremore susceptible to “gaming” since the reward is higher, andsubjects cannot save drafts, raising the possibility of lost datadue to session timeout or connectivity error. We instead as-signed each chart as an individual task. Since the vast ma-jority (95%) of subjects accepted all tasks in sequence, theexperiment adhered to the original within-subjects format.
ResultsTo analyze responses, we replicated Cleveland & McGill’sdata exploration, using their log absolute error measure ofaccuracy: log2(|judged percent - true percent| + 1
8 ). We firstcomputed the midmeans of log absolute errors1 for each chart(Figure 3). The new results are similar (though not identical)to the originals: the rough shape and ranking of judgmenttypes by accuracy (T1-5) are preserved, supporting the valid-ity of the crowdsourced study.
Next we computed the log absolute error means and 95%confidence intervals for each judgment type using bootstrap-ping (c.f., [7]). The ranking of types by accuracy is consistentbetween the two experiments (Figure 4). Types 1 and 2 arecloser in the crowdsourced study; this may be a result of asmaller display mitigating the effect of distance. Types 4 and5 are more accurate than in the original study, but positionencoding still significantly outperformed length encoding.
We also introduced two new judgment types to evaluate an-gle and circular area encodings. Cleveland & McGill con-ducted a separate position-angle experiment; however, theyused a different task format, making it difficult to compare
1The midmean–the mean of the middle two quartiles–is a robust measureless susceptible to outliers. A log scale is used to measure relative propor-tional error and the 1
8 term is included to handle zero-valued differences.
Test % difference in area between elements
�29D. Koop, DSC 201, Fall 2018
[Heer & Bostock, 2010]
Answer: B is ~2.5 larger (in area) than A
Positions
Rectangular areas
(aligned or in a treemap)
Angles
Circular areas
Cleveland & McGill’s Results
Crowdsourced Results
1.0 3 .01.5 2 .52 .0Log Error
1.0 3 .01.5 2 .52 .0Log Error
Results Summary
�30D. Koop, DSC 201, Fall 2018
[Munzner (ill. Maguire) based on Heer & Bostock, 2014]
Visualization Tools• Analysis Apps: Tableau, Excel, SAS • Illustration Apps: Illustrator, Inkscape • R Libraries: base, ggplot • Python Modules: matplotlib, seaborn, bokeh, altair • Lower-level Frameworks: D3, Processing • Many, many more: Google "data visualization tools"
�31D. Koop, DSC 201, Fall 2018
Tableau Overview• Grew out of research at Stanford University on how to explore
multidimensional datasets & relational databases • Tableau Desktop: standalone (free trial, student license) • Tableau Public: cloud-based system (free) • Tableau Vizable: mobile app • Tableau's Introduction Videos • https://www.youtube.com/watch?v=6py0jyZc7K4
�32D. Koop, DSC 201, Fall 2018
Data In Tableau
• Categorical data = Dimension • Quantitative data = Measures
�33D. Koop, DSC 201, Fall 2018
Attributes
Attribute Types
Ordering Direction
Categorical Ordered
Ordinal Quantitative
Sequential Diverging Cyclic
Tableau Example
�34D. Koop, DSC 201, Fall 2018