Learning Perceptual Kernels for Visualization Design Ça ğ atay Demiralp Michael Bernstein Jeffrey Heer Stanford University Stanford University University of Washington Interactive Data Lab @ UW
15 11.25 7.5 3.75 0 13 9.75 6.5 3.25 Visualizations 0 0 4 7 11 14 0 5 10 15 20 Leverage Perception balance in favor England balance against 2
Engineering Perception Into Visualization Design?
A Measure of Perceptual Reality Perceptual Kernel 2D Projection 4
What are Perceptual Kernels Useful For? 5
Automating Visualizations
Palette Design l n 7
Palette Design l original n 2 3 4 5 6 7 8 9 ed reordered 10 8
Palette Design 2D Projection l n 2 3 4 5 6 7 8 9 ed reordered 10 9
Palette Design l original n 2 3 4 5 6 7 8 9 ed reordered 10 10
Palette Design l original n 2 3 4 5 6 7 8 9 ed reordered 10 Palettes re-ordered to maximize perceptual discriminability 11
Visual Embedding: A Model for Visualization 12
Visualizations as Functions Visual Data f : X → Y Primitives Points Y X color quantitative size ordinal shape nominal orientation … texture … 13
Visual Embedding d d Y X y x 1 1 small y 2 x 2 small f : X → Y Visual Data y 3 Primitives Points x 3 large y 4 large x 4 Y X quantitative color ordinal size nominal shape … orientation texture … 14
Visual Embedding d d Y X y x 1 1 small y 2 x 2 small f : X → Y Visual Data y 3 Primitives Points x 3 large y 4 large x 4 Y X quantitative color ordinal size nominal shape … orientation texture … 15
Visual Embedding d d Y X y x 1 1 small y 2 x 2 small f : X → Y Visual Data y 3 Primitives Points x 3 large y 4 large x 4 Y X quantitative color ordinal size nominal shape … orientation texture NOT NEED TO BE METRIC … 16
Rank Correlations - - i f l s Kernel (Tm) A - Color Names B e- CIELAB C CIEDE2000 - D A B C D A 1.00 0.75 0.67 0.59 B 1.00 0.81 0.77 C 1.00 0.87 D 1.00 17
Rank Correlations - - i f l s Kernel (Tm) A - Color Names B e- CIELAB C CIEDE2000 - D A B C D A 1.00 0.75 0.67 0.59 B 1.00 0.81 0.77 C 1.00 0.87 D 1.00 18
Rank Correlations - - i f l s Kernel (Tm) A - Color Names B e- CIELAB C CIEDE2000 - D A B C D A 1.00 0.75 0.67 0.59 B 1.00 0.81 0.77 C 1.00 0.87 D 1.00 19
Rank Correlations - - i f l s Kernel (Tm) A - Color Names B e- CIELAB C CIEDE2000 - D A B C D A 1.00 0.75 0.67 0.59 B 1.00 0.81 0.77 C 1.00 0.87 D 1.00 20
Rank Correlations - - i f l s Kernel (Tm) A - Color Names B e- CIELAB C CIEDE2000 - D A B C D A 1.00 0.75 0.67 0.59 B 1.00 0.81 0.77 C 1.00 0.87 D 1.00 21
Rank Correlations - - i f l s Kernel (Tm) A - Color Names B e- CIELAB C CIEDE2000 - D A B C D A 1.00 0.75 0.67 0.59 B 1.00 0.81 0.77 C 1.00 0.87 D 1.00 22
Cluster Connectivity Encode community clusters in a character co-occurrence graph. 23
24
CONTRIBUTIONS 25
CONTRIBUTIONS 1) Estimate perceptual kernels shape size color shape-size shape-color size-color 26
CONTRIBUTIONS 2) Compare alternative judgment types pairwise-5 pairwise-9 triplet triplet manual matching discrimination 27
CONTRIBUTIONS 3) Assess using existing models vs. I ∼ M β Garner’s Integrality Stevens’ Power Law CIELAB CIEDE2000 Color Names 28
CONTRIBUTIONS 4) Demonstrate in visualization automation designing palettes visual embedding 29
Crowd-sourcing Perceptual Kernels 30
Study Overview Variables s hape Tableau size color Tableau s hape-size shape-color size-color Tasks reference a b c a b pairwise-5 pairwise-9 triplet matching triplet discrimination manual spatial arrangement L5 L9 Tm Td SA Subjects Platform 600 Turkers based in the US 95% approval rate minimum 100 approved HITs 31
Univariate Perceptual Kernels L5 L9 SA Tm Td shape color size 32
Bivariate Perceptual Kernels L5 L9 SA Tm Td shape-color shape-size size-color 33
Judgment Tasks 1.Pairwise rating on 5-point scale (L5) 2.Pairwise rating on 9-point scale (L9) 3.Triplet ranking with matching (Tm) 4.Triplet ranking with discrimination (Td) 5.Spatial arrangement (SA) 34
Judgment Tasks 1. Pairwise rating on 5-point scale (L5) 35
Judgment Tasks 1. Pairwise rating on 5-point scale (L5) 36
Judgment Tasks 2. Pairwise rating on 9-point scale (L9) 37
Judgment Tasks 3. Triplet ranking with matching (Tm) 38
Judgment Tasks 3. Triplet ranking with matching (Tm) 39
Judgment Tasks 4. Triplet ranking with discrimination (Td) 40
Judgment Tasks 5. Spatial arrangement (SA) 41
Judgment Tasks 5. Spatial arrangement (SA) 42
Perceptual Kernels & Models of Perception 43
Size (Tm) perceptual kernel 2D projection Consistent with Stevens’ Power Law! 44
Stevens’ Power Law electric length shock ( β =1.1) ( β =3.5) Perceived Intensity ( I ) brightness ( β =0.5) stimulus dependent I ∼ M β exponent True Magnitude ( M ) 45
Stevens’ Power Law Fit 46
Stevens’ Power Law Fit 47
Stevens’ Power Law Fit 48
CONTRIBUTIONS 3) Assess using existing models vs. I ∼ M β Garner’s Integrality Stevens’ Power Law CIELAB CIEDE2000 Color Names details are in the paper 49
Which Judgment Task to Use? 50
Triplet matching (Tm) reference lowest variance, most robust, shortest unit a b reference a Triplet comparisons (Tm & Td) longest experiment time, highest cost a b b c Pairwise Likert ratings (L5 & L9) faster & cheaper than triplet comparisons Manual spatial arrangement (SA) fastest, cheapest high variance, high sensitivity 51
Triplet matching (Tm) reference best lowest variance, most robust, shortest unit a b reference a Triplet comparisons (Tm & Td) longest experiment time, highest cost a b b c Pairwise Likert ratings (L5 & L9) faster & cheaper than triplet comparisons Manual spatial arrangement (SA) fastest, cheapest high variance, high sensitivity 52
Triplet matching (Tm) reference best lowest variance, most robust, shortest unit a b reference a Triplet comparisons (Tm & Td) longest experiment time, highest cost a b b c Pairwise Likert ratings (L5 & L9) faster & cheaper than triplet comparisons Manual spatial arrangement (SA) fastest, cheapest worst high variance, high sensitivity 53
Perceptual Kernels operational model Use ordinal triplet matching CONCLUSIONS unless prohibited by time & cost Avoid manual spatial arrangement Read the paper 54
Acknowledgments IDL Group Members data & source code https://github.com/uwdata/perceptual-kernels https://github.com/uwdata/visual-embedding 55
Data Processing Pairwise judgments Produce a distance matrix directly Identical pairs to detect spammers Triplet judgments Generalized non-metric multidimensional scaling Use triplets with two identical elements to detect spammers Spatial arrangements Align to a reference and filter-out the outliers Planar Euclidean distances produce a distance matrix 56
Palette Design l n 2 3 4 5 6 7 8 9 ed 10 57
What About Context? 58
What About Context? What about it? 59
What About Context? 60
What About Context? early results suggest no significant effect 61
Why Tableau? 62
Why Tableau? I have Tableau stocks 63
Why Tableau? I have Tableau stocks? 64
Why Tableau? I have Tableau stocks? 65
Why Tableau? Manually designed with perceptual considerations in mind discriminability, saliency and naming of colors, robustness to spatial overlap of shapes Provides ecological validity and good baseline 66
What About Individual Differences? 67
Per-subject SAs: size The layout with gray background is the medoid of the layouts in affine space. 68
Sensitivity shape color size 1.1 1.1 1.1 1 1 1 0.9 0.9 0.9 corr 0.8 0.8 0.8 L5 L5 L5 0.7 L9 0.7 L9 0.7 L9 SA SA SA Tm Tm Tm 0.6 0.6 0.6 Td Td Td 0.5 0.5 0.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 shape-color shape-size size-color 1.1 1.1 1.1 1 1 1 0.9 0.9 0.9 corr 0.8 0.8 0.8 L5 L5 L5 0.7 0.7 0.7 L9 L9 L9 SA SA SA Tm Tm Tm 0.6 0.6 0.6 Td Td Td 0.5 0.5 0.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 69
Why SA Performs Poorly? 70
Why SA Performs Poorly? Unstructured nature, leading to higher variance across subjects Expressivity limited to two dimensions expression of perceptual structures. 71
Why Tm Outperforms Td? 72
Recommend
More recommend
