1. What Makes a Visualization Memorable?
327 cited/ Published in 2013/ IEEE Transactions on Visualization and Computer Graphics
Mishelle A. Borkin, Azalea A. Vo, Zoya Bylinskii, Phillip Isola, Shashank Sunkavalli, Aude Oliva, Hanspeter Pfister
2. 1
Motivation
• Memorability of images of natural scenes is consistent
• Memorability
Engagement
Effectiveness
• Provide a basis for understanding a number of cognitive aspects of visualization
• Build a new broad taxonomy of static visualizations
• Systematically analyzes memorizing intuition
• Not about memorability or comprehensibility
• Memorability of visualization as images
3. 2
Visualization taxonomy
Reference
R. L. Harris. Information graphics: A comprehensive illustrated reference. Oxford University Press, 1999.
Y. Englehardt. The language of graphics: A framework for the analysis of syntax and meaning in maps, charts and
diagrams. PhD thesis, Institute for Logic, Language and Computation, University of Amsterdam, 2002.
To aid characterization of
the visualizations
To gain insight,
Used in experiment
4. 3
Data Collection and Annotation
• Scrap 5,693 data visualizations
• Chosen based on
A large number of static visualizations
Able to scrap
• Infographic and scientific publications
have multiple vis to explain concept
Fully responsible for whole story
• Exclude multiple
• Remain 2,070 single visualizations
5. 4
Analysis of Visualization types
• Scientific Publication
Diagram
Basic visual encodings
• Infographic
Diagram
Table
• News Media & Government
Common forms of basic visual encodings
• Text visualization is not main stream
• Real-world examples
6. 5
Memorability Experiment: Hypotheses
H1) Participants will perform worse (i.e., overall have a harder time remembering visualizations) as
compared to natural images/photos.
H2) A visualization is more memorable if it includes a pictogram or cartoon of a recognizable image.
H3) A visualization is more memorable if there is more color.
H4) A visualization is more memorable if it has low visual density.
H5) A visualization is more memorable if it is more “minimalist” (i.e., “good” data-ink ratio).
H6) A visualization is more memorable if it includes a “familiar” visualization type (i.e., basic graph type
taught in school).
H7) A visualization is less memorable if it comes from a scientific publication venue.
7. 6
Memorability Experiment: Target images
• Select subset of 410 images
• Attributes rankings generated by 3 visualization researchers
Rate individual
If not same, make consensus
• Effect of chart junk
Data-ink ratio good : 145 | bad : 103 | medium 162
• Consider visualization categories as well
8. 7
Memorability Experiment: Participants and Set-up
• Target images are repeated twice
• Press a key at a second time
• Total 410 images of targets
• Over 17 levels
Each with 120 images comprised of targets and filters
Each takes 4.8 minute
• 1 sec for each image | 1.4 sec for gap
• 91-109 images apart
• Participants filtering
Vigilance repeats for twice with spacing of 1-7 images
False-alarm on > 50 % of last 30 non-repeat filter image
To begin, should pass practice trial miss rate < 50 % & false alarm rate < 30 %
9. 8
Memorability Experiment: Design & Analysis
• Based on signal detection theory
• Hit rate (HR) : 𝐻𝑅 =
𝐻𝐼𝑇𝑆
𝐻𝐼𝑇𝑆+𝑀𝐼𝑆𝑆𝐸𝑆
Response on the second presentation
• False alarm rate (FAR) : 𝐹𝐴𝑅 =
𝐹𝐴
𝐹𝐴+𝐶𝑅
CR: number of correct rejection
Response on the first presentation
To distinguish confusion by familiar image
• d-prime metric (sensitivity index) : 𝑑′ = 𝑍 𝐻𝑅 − 𝑍 𝐹𝐴𝑅
Z : inverse cumulative Gaussian distribution
Memorability score
Confused for previous memory (low FAR) will result low score
10. 9
Memorability Experiment: Result & Discussion (H1)
• Visualization : HR of 55.36 % | FAR of 13.17%
• Scene : HR of 67.5 % | FAR of 10.7%
• Face : HR of 53.6 % | FAR of 14.5%
Support H1
• Spearman’s rank correlations of 0.83 for HR, 0.78 for
FAR, and 0.81 for d-prime
Participants split into two group
• Consistency of memorability score is verified
H1) Participants will perform worse (i.e., overall have a harder time
remembering visualizations) as compared to natural
images/photos.
11. 10
Memorability Experiment: Result & Discussion (H2)
• 145 contain either photographs, cartoons,
or human recognizable object = pictogram
• With pictogram: score (M) = 1.93
• Without pictogram: score (M) = 1.14
Support H2
Overall Without pictogramH2) A visualization is more memorable if it includes a
pictogram or cartoon of a recognizable image.
12. 11
Memorability Experiment: Result & Discussion (H3 & H4)
H3) A visualization is more memorable if there is more color.
• Color ≥ 7 : score (M) = 1.71
• 2 ≤ Color ≤ 6 : score (M) = 1.48
• Color = 1 : score (M) = 1.18
Support H3
• Visual density = High : score (M) = 1.83
• Visual density = Low : score (M) = 1.28
Refute H4
H4) A visualization is more memorable if it has low visual density.
13. 12
Memorability Experiment: Result & Discussion (H5 & H7)
H5) A visualization is more memorable if it is more “minimalist” (i.e.,
“good” data-ink ratio
• Data-ink ratio
• Bad : score (M) = 1.81
• Good : score (M) = 1.23
Refute H5
• Infographic : score (M) = 1.99
• Scientific Publication : score (M) = 1.48
• News Media : score (M) = 1.46
• Government : score (M) = 0.86
Refute H7
H7) A visualization is less memorable if it comes from a scientific
publication venue.
14. 13
Memorability Experiment: Result & Discussion (H6)
• Grid/matrix, trees and networks and diagram have the highest memorability scores
• High FAR and low HR in bar vis
• Possible interpretation
Low frequency contribute to uniqueness
Image memorability correlate with recognizing natural things
Refute H6
H6) A visualization is more memorable if it includes a “familiar” visualization type (i.e., basic graph type taught in school).
15. 14
Conclusion
• Visualizations are intrinsically memorable with consistency across people
Natural scenes
Colorful
Inclusion of a human recognizable object enhance memorability
Low data-to-ink ratios
High visual densities (i.e., more chart junk and “clutter”)
Natural looking visualization
• Future work to examine subtypes
Editor's Notes
기존에도 비슷한 연구가 있긴 했으나
특정 그래프를 비교하는 페이퍼 였음
차이점은 target a wide variety of visualization
Automatic 하게 할 수 있을 만큼 충분한 양의 데이터가 있는지를 기준으로 임의로 선택
Amazon’s mechanical Turk
스피어만 상관 계수 : 데이터에 순위를 매겨 상관 계수 구함
1에 가까울 수록 두 변수간의 상관 관계가 일치
Chance 는 랜덤 변수
This memorability study examined the memorability of visualizations as if they were images and not memorability based on engagement and comprehension of the visualization.
H7) Visual.ly 에서 자료 수집 사전 검열된 자료인 만큼 fancy 한 looking 을 가지고 있기 때문에 다른 것에 비해 bias 가 있을 수 있는 가능성
Government source 는 aesthetic 심미적 요소를 고려하지 않음