evaluating perceptual bias during geometric scaling of scatterplots percept… · evaluating...
TRANSCRIPT
Evaluating Perceptual Bias During Geometric Scaling of ScatterplotsYating Wei1, Honghui Mei1, Ying Zhao2, Shuyue Zhou1, Bingru Lin1, Haojing Jiang2, Wei Chen1
1: State Key Lab of CAD&CG, Zhejiang University2: Central South University
Introduction
• Scatterplots are a common type of visualization.• Various display devices• Collaborative data analysis
Introduction
Motivation
• Geometric scaling• Enlarge or shrink the entire plot
together with the objects inside it synchronously and proportionally.
Geometric scaling may cause a perceptual bias.
The deviation between the perceived and physical values of visual features.
• Design decisions of scatterplot
Introduction
[Wang et al. 2019]
[Tory et al. 2009]
[Heer et al. 2006]
…
Related Work
• Usage issues of scatterplots in multiple devices
Introduction
[Butscher et al. 2018]
[Blascheck et al. 2019]
[Ni et al. 2006]
Related Work
Introduction
1Conduct evaluation experiments to verify the existence of the bias.
2Obtain instructive findings to guide visual design.
Goal
Hypotheses
• H3: Changing the radius of points can reduce the bias.
• H1: Geometric scaling causes a bias in the perceived visual features of scatterplots.
• H2: The bias caused by geometric scaling can be affected by data distribution.
Numerosity Correlation Cluster separation
Normal
Uniform
radii
Hypotheses
• H3: Changing the radius of points can reduce the bias.
• H1: Geometric scaling causes a bias in the perceived visual features of scatterplots.
• H2: The bias caused by geometric scaling can be affected by data distribution.
Numerosity Correlation Cluster separation
Normal
Uniform
radii
Hypotheses
• H3: Changing the radius of points can reduce the bias.
• H1: Geometric scaling causes a bias in the perceived visual features of scatterplots.
• H2: The bias caused by geometric scaling can be affected by data distribution.
Numerosity Correlation Cluster separation
Normal
Uniform
radii
Experimental Design
2D scatterplot- Aspect ratio = 1- 𝑥𝑥,𝑦𝑦 ∈ [0,1]- Gray background & black thin borders- Black circles with the same radii
Experimental Design
StimuliCommon displays
25% 40% 63% 100% 159% 252% 400%
originshrink enlarge
Seven Scale Ratios
• The value range of the testing visual feature
−6 level −5 level −4 level +6 level+5 level+4 level
Experimental Design
Variable and Data
Numerosity𝑛𝑛 ∈ [11,730]
Correlation𝑟𝑟 ∈ [0.15,0.9]
Cluster Separation𝑠𝑠 ∈ [0.11,0.65]
• Simulate scatterplot scaling scenarios.• Measure people’s subjective experience.
Experimental Design
2IFC & 2WS Method
• 2IFC (two-interval forced choice)• Measure the subjective experience of a person through his/her choice pattern.• 2IFC sequence: measure the bias from the choice pattern.
• 2WS (two-way staircase)• Decide the order of appearance of feature levels
Make the right choiceMake the right choice
Make the wrong choice
Experimental Design
2IFC & 2WS Method
Experimental Design
Experimental settings
E1(H1) E2(H2) E3(H3)
VariableScale ratio, feature, feature level
Scale ratio, feature, feature level
Point radius, feature, feature level
ConstantPoint radius Point radius Two scale ratios
Data distributionNormal distribution Uniform
distributionNormal distribution
63% 252%
1. The overall difficulty level (DL) of making 2IFC judgments
Do the changes of scatterplot size make your judgments on the perception of normal-distributed numerosity difficult? If so, then how difficult is it?
2. The difficulty tendency (DT) of making 2IFC judgments with different scale ratios or point radii.
How does the difficulty of making judgments change on the perception of normal-distributed numerosity with the changes in scatterplot size from small to large?
Experimental Design
Subjective Questionnaire• Two types of questions
Pilot study Preparation Introduction & tutorial Pre-experiment Formal experiment
10 2 3 4
• Participants • Apparatus• 20 participants • an independent, quiet laboratory
• a Dell 25 in U2515H monitor
Experimental Design
Procedure
EXPERIMENTAL RESULTS
• Abnormal choice patterns identification.• Fiercely fluctuating• Early fluctuating
Experimental Results
Analysis ApproachFe
atur
e Le
vels
Trials
Feat
ure
Leve
ls
Trials
• Bias measurement• PSE (point of subjective equality) − POE (point of objective equality)
• Significance analysis• Shapiro−Wilk test• Non-parametric Friedman• ANOVA test
Experimental Results
Analysis Approach
Feature value of test scatterplotPr
ob. T
est<
Refe
renc
e
H1: Geometric scaling causes a bias in perceiving the three visual features. The bias has a significant linear relationship with the scale ratio.
FULLY CONFIRMED!
Experimental Results
Objective Result Analysis —— H1
Scale Ratio (%)
Bias
Numerosity
65 → 89
Prob
. Tes
t<Re
fere
nce
Feature value of test scatterplot
Experimental Results
Objective Result Analysis —— H1Numerosity
Bias
Scale Ratio
Perceived Numerosity
Experimental Results
Objective Result Analysis —— H1
Scale Ratio Perceived Correlation Perceived Cluster Separation
Correlation Cluster Separation
Bias
Bias
Scale Ratio (%) Scale Ratio (%)
Experimental Results
Objective Result Analysis —— H2
Scale Ratio (%) Scale Ratio (%) Scale Ratio (%)
Bias Bias
Bias
Correlation Cluster SeparationNumerosity
H2: The bias can be affected by data distribution. FULLY NEGATED!
H3: Changing the point radius can reduce the bias. PARTIALLY CONFIRMED!
Experimental Results
Objective Result Analysis —— H3
Numerosity
Bias
Point Radius
E3 (6
3%)
Numerosity
Point Radius
Bias
E3 (2
52%
)
Experimental Results
Objective Result Analysis —— H3Correlation Cluster SeparationNumerosity
E3 (6
3%)
E3 (2
52%
)
Experimental Results
Objective Result Analysis —— H3Correlation Cluster SeparationNumerosity
E3 (6
3%)
E3 (2
52%
)
Experimental Results
Objective Result Analysis —— H3Correlation Cluster SeparationNumerosity
E3 (6
3%)
E3 (2
52%
)
Experimental Results
Subjective Result Analysis
Experimental Results
Summary• Geometric scaling causes a bias on the perception of the three visual
features.
• No significant difference occurs between the biases measured from normally and uniformly distributed scatterplots.
• Changing the point radius can correct the bias.
• Limited experimental content• High-level perceptions of scatterplot• Diverse data distributions• More visual channels• Different display devices and screen resolutions
• No quantitative model• To describe the bias• To guide the bias correction
Limitation
1Evaluation Experiment
2Instructive findings
Conclusion
Thank you!
This work is supported by:The National Key Research and Development Program of China (No. 2018YFB1700403) The National Natural Science Foundation of China (No. 61772456, 61761136020, U1609217, 61672538 and 61872388).
Yating Wei, Honghui Mei, Ying Zhao, Shuyue Zhou, Bingru Lin, Haojing Jiang, Wei Chen
Evaluating Perceptual Bias During Geometric Scaling of Scatterplots