User Centered Visualization

For the Boston DataVis Meetup @ MSR NERD
http://www.meetup.com/bostondatavis/events/202511862/

Transcript

1. User-Centered Visualization
2. I’m Lane (@laneharrison — postdoc)
5. valt.cs.tufts.edu
6. User-Centered? What does it mean to be
7. User What does it mean to be
8. Perceptual
 Ability
9. Perceptual
 Ability Cognitive
 State
10. Perceptual
 Ability Cognitive
 State Cognitive
 Ability
11. 3 user-parts 3 experiments
12. 1. Leveraging Perception 2. Influencing the User 3. Individual Differences
13. Leveraging perception
14. A brief history of vis (research)
15. 1990: VIS is formed
16. 1990-2000: The Techniques Era
19. 2000-2010: The Techniques Era, part 2
22. With 20 years of techniques,
 how do we choose between them?
23. “How do I know when to bin a scatterplot?”
28. just-noticeable difference (jnd)
29. imagine yourself in a dark room…
34. But you don’t have to take my word for it
36. @jnd1er
37. In the field of psychophysics, that branch of experimental psychology that studies sensation and perception, a jnd is the amount that something must be changed for the difference to be noticeable, defined to mean that the change is detectable half the time. My goal is to make a noticeable difference -- many jnds worth -- in human-centered technology. I started my career as an experimental/mathematical psychologist in psychophysics, and my love of the exquisite sensitivity and dynamic range of hearing and seeing, as well as the power of human perception has stayed with me. - Don Norman (@jnd1er)
38. back to scatterplots
39. after hours of torturing psych students…
40. 0. 0.00 0.06 0.12 0.18 0.24 ;ĂͿƐĐĂƩĞƌƉůŽƚ͕ZĞŶƐŝŶŬ r 0.0 0.2 0.4 0.6 0.8 1.0 0.00 0.06 0.12 0.18 0.24 ƽ ƽ ƽ ƽ ƽ ƽ ƽ භ from above from below JND Rensink plotted JND as a function of correlation (r)
41. The lower the r, 
 the higher the jnd
42. 0.00 0.06 0.12 0.18 0.24 ;ĂͿƐĐĂƩĞƌƉůŽƚ͕ZĞŶƐŝŶŬ r 0.0 0.2 0.4 0.6 0.8 1.0 0.00 0.06 0.12 0.18 0.24 ƽ ƽ ƽ ƽ ƽ ƽ ƽ භ from above from below JND To see a difference in data with correlation of 0.3, the comparison r must be +/- 0.2.
43. increase to 0.5! To see a diff,
44. decrease to 0.2! To see a diff,
45. Rensink’s insight…
46. ;ďͿƐĐĂƩĞƌƉůŽƚ r 0.0 0.2 0.4 0.6 0.8 1.0 0.00 0.06 0.12 0.18 0.24 ƽ ƽ ƽ ƽ ƽ ƽ ;ĐͿƐĐĂƩĞƌƉůŽƚƌĞŐƌĞƐƐŝŽŶ rA 0.0 0.2 0.4 0.6 0.8 1.0 0.00 0.06 0.12 0.18 0.24 ƽ ƽ ƽ ƽ ƽ ƽ 0.0 0.00 0.06 0.12 0.18 0.24 This trend is roughly linear! 0.00 0.06 0.12 0.18 0.24 ;ĂͿƐĐĂƩĞƌƉůŽƚ͕ZĞŶƐŝŶŬ r 0.0 0.2 0.4 0.6 0.8 1.0 0.00 0.06 0.12 0.18 0.24 ƽ ƽ ƽ ƽ ƽ ƽ ƽ භ from above from below JND
48. r=1.0 r=0.0
49. Weber’s Law dp = k * d(r) r __
50. E. H. Weber
51. Lane’s insight…
52. ;ďͿƐĐĂƩĞƌƉůŽƚ r 0.0 0.2 0.4 0.6 0.8 1.0 0.00 0.06 0.12 0.18 0.24 ƽ ƽ ƽ ƽ ƽ ƽ ;ĐͿƐĐĂƩĞƌƉůŽƚƌĞŐƌĞƐƐŝŽŶ rA 0.0 0.2 0.4 0.6 0.8 1.0 0.00 0.06 0.12 0.18 0.24 ƽ ƽ ƽ ƽ ƽ ƽ 0.0 0.00 0.06 0.12 0.18 0.24 ;ĂͿƐĐĂƩĞƌƉůŽƚ͕ZĞŶƐŝŶŬ r 0.0 0.2 0.4 0.6 0.8 1.0 0.00 0.06 0.12 0.18 0.24 ƽ ƽ ƽ ƽ ƽ ƽ ƽ භ from above from below JND If scatterplots follow Weber’s law like this…
53. ;ďͿƐĐĂƩĞƌƉůŽƚ r 0.0 0.2 0.4 0.6 0.8 1.0 0.00 0.06 0.12 0.18 0.24 ƽ ƽ ƽ ƽ ƽ ƽ ;ĐͿƐĐĂƩĞƌƉůŽƚƌĞŐƌĞƐƐŝŽŶ rA 0.0 0.2 0.4 0.6 0.8 1.0 0.00 0.06 0.12 0.18 0.24 ƽ ƽ ƽ ƽ ƽ ƽ 0.0 0.00 0.06 0.12 0.18 0.24 ;ĂͿƐĐĂƩĞƌƉůŽƚ͕ZĞŶƐŝŶŬ r 0.0 0.2 0.4 0.6 0.8 1.0 0.00 0.06 0.12 0.18 0.24 ƽ ƽ ƽ ƽ ƽ ƽ ƽ භ from above from below JND What if different charts follow Weber’s law differently?
55. r = -1 r = -0.8 r = -0.3 r = 0.3 r = 0.8 r = 1 scatterplot parallel coordinates (pcp) stackedarea r = -1 r = -0.8 r = -0.3 r = 0.3 r = 0.8 r = 1 scatterplot parallel coordinates (pcp) stackedarea
56. coordinates (pcp) stackedarea stackedline stackedbar r = -1 r = -0.8 r = -0.3 r = 0.3 r = 0.8 r = 1 scatterplot parallel coordinates (pcp) stackedarea
57. stackedbar donut radar line r = -1 r = -0.8 r = -0.3 r = 0.3 r = 0.8 r = 1 scatterplot parallel coordinates (pcp) stackedarea
58. radar line ordered line r = -1 r = -0.8 r = -0.3 r = 0.3 r = 0.8 r = 1 scatterplot parallel coordinates (pcp) stackedarea
59. How can we test all these visualizations?
60. A: Crowdsourcing
61. n=1687 on AMT (Amazon’s Mechanical Turk)
62. Results
63. scatterplot − positive r 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 • • • • • • • • • • • • • scatterplot − negative r jnd 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 • • • • • • • • • • • • • parallel coordinates − positive r jnd 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 • • • • • • • • • • • • • parallel coordinates − negative r jnd 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 • • • • • • • • • • • • • stackedarea − positive r jnd 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.0 0.1 0.2 0.3 0.4 0.5 0.6 • • • • • • • • • • • • • stackedline − positive r 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 • • • • • • • • • • • • stackedline − negative r jnd 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 • • • • • • • • • • • • stackedbar − positive r jnd 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 • • • • • • • • • • • • stackedbar − negative r jnd 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 • • • • • • • • • • • • donut − positive r jnd 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.0 0.1 0.2 0.3 0.4 0.5 0.6 • • • • • • • • • • • • line − positive 0.3 0.4 0.5 0.6 • • • • • • • line − negative jnd 0.3 0.4 0.5 0.6 • • • • • • • • • • radar − positive jnd 0.3 0.4 0.5 0.6 • • • • radar − negative jnd 0.3 0.4 0.5 0.6 • • • • • • • • • • • ordered line − positive jnd 0.3 0.4 0.5 0.6 Lots of Plots
65. All follow 
 Weber’s Law!
66. Findings
67. Not all perceptual spaces are symmetric.
68. 0.7 0.8 0.85 scatterplot (positive) scatterplot (negative) parallel coordinates (negative) parallel coordinates (positive) Symmetric Not Symmetric
69. sp and pcp 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 parallelCoordinates scatterplot positive negative -PCP as good as scatterplots +PCP terrible
70. To show correlations accurately in parallel coordinates plots, flip the axes to show as many negative correlations as possible. Design guideline: scatterplot (positive) scatterplot (negative) parallel coordinates (negative) parallel coordinates (positive) scatterplot (positive) scatterplot (negative) parallel coordinates (negative)
71. Which stacked-variant 
 wins?
72. stackedarea, stackedline and stackedbar 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 stackedline stackedarea stackedbar positive negative parallel coordinates (pcp) stackedarea stackedline scatterplot parallel coordinates (pcp) stackedarea stackedline stackedbar donut radar line
73. stackedarea, stackedline and stackedbar 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 stackedline stackedarea stackedbar positive negative parallel coordinates (pcp) stackedarea stackedline scatterplot parallel coordinates (pcp) stackedarea stackedline stackedbar donut radar line More detail More abstract
74. Design guideline: Less isn’t always more. Sometimes less “minimal” charts also yield less ambiguity.
75. Radar versus Line?
76. line and radar 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 line radar positive negative
77. line and radar 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 line radar positive negative Longer line Shorter line
78. Design guideline: If you have to use a line chart, make sure it’s wide enough.
79. Perceptually-backed ranking of accuracy
80. model fit results 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 ordered line line radar stackedline stackedarea stackedbar donut scatterplot parallel coordinates scatterplot, Rensink positive negative JND rA Take the models, obtain rankings at each |r|
82. The paper: “Ranking Visualizations of Correlation Using Weber's Law.” Lane Harrison, Fumeng Yang, Steven Franconeri, Remco Chang . To appear, InfoVIS 2014 (pdf online)
83. (I still don’t know when to bin a scatterplot)
84. Influencing The User
85. Make Decisions Reason Perceive D A T A
86. Make Decisions Reason Perceive D A T A Visualization is also used in more serious contexts:
87. What do we know about affect?
88. Affect influences creativity. Affect alters how we decide under uncertainty. Affect modulates visual perception. (Lewis et al., 2011) (Fredrickson, et al., 2005) (Vuilleumier et al., 2007)
89. How does Emotion impact graphical perception?
90. leverage well- studied graphical perception task adapt well-studied emotion-priming techniques Experiment combining emotion and graphical perception. Let`s start with what we know...
91. Position Angle > (Cleveland & McGill,1984) a. Which of the two is larger? b. What percentage is the smaller of the larger?
92. Cleveland & McGill's Results 1.0 1.5 2.0 2.5 3.0 T1 T2 T3 T4 T5 Log Error Crowdsourced Results T1 T2 Useful for visualization design guidelines; heavily replicated. (image: Heer and Bostock, 2009)
93. 1.0 1.5 2.0 2.5 3.0 T5 Log Error Crowdsourced Results 1.0 1.5 2.0 2.5 3.0 T1 T2 T3 T4 T5 T6 T7 T8 T9 Log Error (image: Heer and Bostock, 2009) Validated AMT for graphical perception studies. Heer & Bostock, 2009
94. A B 100 0 A B 100 0 A B 100 0 A B A B tree A B A B 100 0 A B V1 V2 V3 V4 V5 V6 V7 V8 bar chart (adjacent comparison) bar chart (non-adjacent comparison) horizontal bar chart (adjacent comparison) stacked bar chart (adjacent comparison) pie chart (ordered) pie chart (unordered) bubble chart treemap
95. A B 100 0 A B 100 0 A B 100 0 A B A B tree A B A B 100 0 A B V1 V2 V3 V4 V5 V6 V7 V8 bar chart (adjacent comparison) bar chart (non-adjacent comparison) horizontal bar chart (adjacent comparison) stacked bar chart (adjacent comparison) pie chart (ordered) pie chart (unordered) bubble chart treemap Replicate Cleveland & McGill! Again!
96. leverage well- studied graphical perception task adapt well-studied emotion-priming techniques Experiment combining emotion and graphical perception.
97. How do we induce Emotion? Stories Visual memory != verbal memory Commonly used in web- based studies. (Goeritz., 2007) (Shackman et al., 2006)
98. How do we induce Emotion? Stories Visual memory != verbal memory Commonly used in web- based studies. (Goeritz., 2007) (Shackman et al., 2006) An excerpt...
99. No one could say for sure how long she would live, but continued hospital care was clearly pointless. Nor could she go home: she needed more attention than her family could provide... The problem was, she had no place to go. There was a hospice facility near her house, but it would accept her only if she would die within six days... - Excerpt from Looking for a Place to Die, Theresa Brown
100. How do we quantify Emotion? How do we induce Emotion? Stories Visual memory != verbal memory Commonly used in web- based studies. (Goeritz., 2007) (Shackman et al., 2006) (Lang et al., 2008, Lewis et al., 2011) 9-point 
 Self-Assessment Manikin (SAM)
101. Pilot 2: how effective is the priming? Pilot 1: is the stimuli valid? Full-study: all 8 chart types Study components:
102. Full-study: 8 chart types Purpose: test whether affect influences graphical perception. Design: 
 - n = 963 on Mechanical Turk
 - 1 random prime, 1 random chart
 - 5 perception tasks per participant
 - between subjects
103. Experiment procedure (Full study & Pilot 2) Pre-Valence Pre-Arousal Post-Valence Post-Arousal Accuracy Verification Question Measure Emotion Random Priming The patient was a fairly young woman and she'd had cancer for as long as her youngest child had been alive... During this past year I've had three instances of car trouble: a blowout on a freeway, a bunch of blown fuses and an out-of-gas situation... A B 100 0 A B 100 0 A B 100 0 A B A B tree A B A B 100 0 A B Random Chart V1 V2 V3 V4 V5 V6 V7 V8 Tasks Which of the two (A or B) is SMALLER? What percentage is the SMALLER of the LARGER? Measure Emotion
104. Full-study: 8 chart types Cleanup: 
 - 299 removed for junk answers
 - n = 664 total...
 - n = 207 successfully primed
105. Full-study: 8 chart types Cleanup: 
 - 299 removed for junk answers
 - n = 664 total...
 - n = 207 successfully primed Two cases: by priming group (664) and by SAM-change (207)...
106. Mean of all participants, regardless of final SAM (n = 664): 0 1 2 3 4 Positively Primed Negatively Primed Means of All Participants No significant difference in error: 
 t(662) = 1.8318; p = .067
107. All participants, regardless of final SAM (n = 664): A B 100 0 A B 100 0 A B 100 0 A B A B tree A B A B 100 0 A B V1 V2 V3 V4 V5 V6 V7 V8 0 1 2 3 4 Positively Primed Negatively Primed
108. Means of primed participants (n = 207): 0 1 2 3 4 Positively Primed Negatively Primed Means of Primed Participants Significant difference in error: 
 t(205) = 3.1560; p = .0018
109. Primed participants (n = 207): A B 100 0 A B 100 0 A B 100 0 A B A B tree A B A B 100 0 A B V1 V2 V3 V4 V5 V6 V7 V8 0 1 2 3 4 Positively Primed Negatively Primed
110. Expert Discussion: 
 Steven Franconeri, Northwestern Positive moods can expand the scope of the perceptual spotlight of attention. Encourage an observer to process a larger spatial area of the world in a single glance. Negative or anxious moods can constrict this spatial area. (Eriksen & St. James, 1986) (Gasper & Clore, 2002; Rowe et al., 2007) (Eysenck & Calvo, 1992) To summarize...
111. Emotion plays an important role in visualization. Emotion influences graphical perception accuracy.
112. Design guideline: When designing for stressful situations, put important information close together.
113. The paper: “Influencing Visual Judgment through Affective Priming.” Lane Harrison, Drew Skau, Steven Franconeri, 
 Aidong Lu, Remco Chang. ACM CHI 2013, (pdf online)
114. Individual Differences
115. Bayesian Reasoning The probability that a woman over age 40 has breast cancer is 1%. However, the probability that mammography accurately detects the disease is 80% with a false positive rate of 9.6%. If a 40-year old woman tests positive in a mammography exam, what is the probability that she indeed has breast cancer?
116. Answer Bayes’ theorem states that P(A|B) = P(B|A) * P(A) / P(B). In this case, A is having breast cancer, B is testing positive with mammography. P(A|B) is the probability of a person having breast cancer given that the person is tested positive with mammography. P(B|A) is given as 80%, or 0.8, P(A) is given as 1%, or 0.01. P(B) is not explicitly stated, but can be computed as P(B,A)+P(B,˜A), or the probability of testing positive and the patient having cancer plus the probability of testing positive and the patient not having cancer. Since P(B,A) is equal 0.8*0.01 = 0.008, and P(B,˜A) is 0.093 * (1-0.01) = 0.09207, P(B) can be computed as 0.008+0.09207 = 0.1007. Finally, P(A|B) is therefore 0.8 * 0.01 / 0.1007, which is equal to 0.07944.
117. Visual aids don’t help.
119. What we tested…
121. Original “Full” Structured
122. “full” text
123. structured text
125. Storyboarding
129. n=377 on AMT
130. Results
131. Overall Accuracy
132. But we also tested…
133. 1. Numeracy 2. Locus of Control 3. Spatial Ability
134. 1. Numeracy 2. Locus of Control 3. Spatial Ability
136. Split on Spatial Ability
137. Design Guideline: Pay attention to VIS research. Individual differences matter, but we’re still figuring it out!
139. User- Centered? What does it mean to be
140. Perceptual
 Ability
141. Perceptual
 Ability Cognitive
 State
142. Perceptual
 Ability Cognitive
 State Cognitive
 Ability
143. Perceptual
 Ability Cognitive
 State Cognitive
 Ability User-Centered Visualization
144. What becomes possible?
145. Big 
 Data small computer
146. Big 
 Data small computer Can we halt computation when differences are imperceptible? Perceptual
 Ability
147. Big 
 Data small computer Can we influence the user to analyze more optimally? Cognitive
 State
148. Big 
 Data small computer Can we adapt to user abilities? Cognitive
 Ability
149. Perceptual
 Ability Cognitive
 State Cognitive
 Ability User-Centered Visualization