Marks + Channels

2017 Czech-Austrian summer school Thomas Torsney-Weir © Munzner/Möller Overview

• Marks + channels • Channel effectiveness – Accuracy – Discriminability – Separability – Popout • Channel characteristics – Spatial position – Colour – Size – Tilt (angle) – Shape (glyph) – Stipple (texture) – Curvature

– Motion © Munzner/Möller 2 Readings

• Munzner, “Visualization Analysis and Design”: – Chapter 5 (Marks and Channels) • Colin Ware: – Chapter 4 (Color) – Chapter 5 (Visual Attention and Information that Pops Out) • The Visualization Handbook: – Chapter 1 (Overview of Visualization) • Additional (background) reading – J. Mackinlay: Automating the design of graphical presentations of relational information. ACM ToG,

5(2), 110-141, 1986© Munzner/Möller 3 Marks + Channels

Rectangle Line

color Width

Height

y-position

Text x-position Marks + Channels

• Mark: basic graphical element / geometric primitive: – point (0D) – line (1D) – area (2D) – volume (3D) • Channel: control appearance (of a mark) – position – size – shape – orientation – hue, saturation, lightness – etc. © Munzner/Möller 7 Height: bin size

x-position: bin name Image

Visual Language is a Sign System • Image perceived as a set of signs • Sender encodes information in signs • Receiver decodes information from signs Visual language is a sign system Images perceived as a set of signs • Jacques Bertin Sender encodes information in signs Receiver decodes information from signs – French cartographer [1918-2010]Jacques Bertin – Semiology of Graphics [1967] Semiology of Graphics, 1983 – Theoretical principles for visual encodings

13 According to Bertin ... Marks Points Lines Areas

Position Size

(Grey)Value

Texture

Channels Color Orientation Shape Semiology of Graphics [J. Bertin, 67]

“Polaris: A System for Query, Analysis and Visualization of Multi-dimensional Relational Databases”, Chris Stolte and Pat Hanrahan

What

Where/ how much

• What: categorical – shape – spatial region – colour (hue) • How Much: ordered (ordinal, quantitative) – length (1D) – area (2D) – volume (3D) – tilt – position

– colour (lightness)© Munzner/Möller 17 Mark types • tables: item = point • network: node+link • link types: – connection: relationship btw. two nodes – containment: hierarchy

Marks as Items/Nodes Points Lines Areas

Marks as Links Containment Connection

• expressiveness principle: – visual encoding should express all of, and only, the information in the dataset attributes – lie factor

• effectiveness principle: – importance of the attribute should match the salience of the channel – data-ink ratio

• Effectiveness according to neurophysiology • Cells in Visual Areas 1 and 2 differentially tuned to each of the following properties: – Orientation and size (with luminance) – Color (two types of signal) – Stereoscopic depth – Motion

• perceptual judgement vs. stimulus • Weber’s law: S = In

Magnitude Channels: Ordered Attributes Identity Channels: Categorical Attributes Position on common scale Spatial region

Position on unaligned scale Color hue

Length (1D size) Motion

Tilt/angle Shape

Area (2D size)

Depth (3D position)

Color luminance

Color saturation

Curvature

Volume (3D size) Effectiveness -- Discriminability

• how many colors can I tell apart? • how many levels of grey etc. • Ex: line width

• separable vs. integral channels

• Integral display dimensions – Two or more attributes perceived holistically • Separable dimensions – Separate judgments about each graphical dimension • Simplistic classiﬁcation, with a large number of exceptions and asymmetries [C. Ware, Information Visualization] [C. Ware,

More separable coding pairs © Munzner/Möller 26 Popout - Preattentive processing • parallel (visual processing)

• Marks + channels • Channel effectiveness • Channel characteristics – Spatial position – Color – Size – Tilt (angle) – Shape (glyph) – Stipple (texture) – Curvature

– Motion © Munzner/Möller 32 Channels

• Spatial position: most effective for all data types (remember the power of the plane) • Size: ‘how much’, interacts with others • Shape/Glyph: ‘what channel’ • Stipple/texture: less popular today • Curvature • Motion: large popout effect

2.05D

© Munzner/Möller 35 Thanks to Moritz Wustinger Thanks to Moritz Wustinger Thanks to Moritz Wustinger Thanks to Moritz Wustinger Smiley based on http://upload.wikimedia.org/wikipedia/commons/b/bd/A_Smiley.jpg Color deﬁciency Source: M. Stone Model “Color blindness”

• Flaw in opponent processing – Red-green common (deuteranope, protanope) – Blue-yellow possible (tritanope -- most common) – Luminance channel almost “normal” • 8% of all men, 0.5% of all women • Effect is 2D color vision model – Flatten color space – Can be simulated (Brettel et. al.) – http://colorﬁlter.wickline.org – http://www.colblindor.com/coblis-color- blindness-simulator/ © Munzner/Möller 41 Source: M. Stone Color Blindness

Protanope Deuteranope Tritanope No L cones No M cones No S cones Red / green Blue / Yellow deﬁciencies deﬁciency

Normal Protanope Deuteranope Lightness

• Limited distinguishability (8-14) – Best with Hue – Best choices from Ware:

• Primary color terms are remarkably consistent across cultures [Berlin & Kay, 69]

• Only a small number of colors can be used effectively as categorical labels

• Keep the number of colors for categorical data to less than eight, and use quiet medium grey backgrounds

Tableau Software © Munzner/Möller 51 Large Areas

1 = Red is bigger 2 = Green is bigger 3 = Both look the same

Cleveland & McGill, “A Color-Caused Optical Illusion on a Statistical© Munzner/Möller Graph”, 1983 54 Color Size Illusion

Cleveland & McGill, “A Color-Caused Optical Illusion on a Statistical© Munzner/Möller Graph”, 1983 55 Source: Ware, “Visual Thinking for Design” Maps

• Color in small regions is difﬁcult to perceive, and bright colors in large areas appear bigger

• Use bright, saturated colors for small regions, and use low saturation pastel colors for large regions and backgrounds

Brewer Scales

Nominal

Ordinal

Cynthia Brewer, Color Use Guidelines for Data Representation

• Lightness and saturation are effective for ordinal data because they have an implicit perceptual ordering

• Show ordinal data with a discrete set of color values that change in lightness or saturation

Rogowitz and Treinish, Why should engineers and scientists be worried about color?

• Hue is used to show ordinal data • Not perceptually linear: Equal steps in the continuous range are not perceived as equal steps • Not good for colorblind people

• Learned order • Visually segmented – Solution — isoluminant rainbow – Solution — discretize colormap

C. Ware, “Visual Thinking for Design”

Density Map Colormaps Density Map

Lightness scale

Lightness scale Lightness scale with hue and chroma variation Lightness scale Lightness scale Hue scale with with hue and lightness variation chroma variation Lightness scale Hue scale with © Munzner/Möllerwith hue and Afterlightness slide from M. Stone variation71 chroma variation Hue scale with lightness variation Take-home message

• Quantitative data can be shown with a discrete or continuous colormap

• Use colormaps with a limited hue palette and redundantly vary lightness and saturation, and use discrete colormaps for accuracy

zInappropriate use of colour can be disasterous to the application

• Poorly designed color is confusing – Creates visual clutter – Misdirects attention • Poor design devalues the information – Visual sophistication – Evolution of document and web design • Don Norman: “Attractive things work better”

• Marks + channels • Channel effectiveness • Channel characteristics – Spatial position – Color • Other channels: – Size – Tilt (angle) – Shape (glyph) – Stipple (texture) – Curvature – Motion © Munzner/Möller 76 Channels: Expressiveness Types and Efectiveness Ranks

Magnitude Channels: Ordered Attributes Identity Channels: Categorical Attributes Position on common scale Spatial region

Position on unaligned scale Color hue

Length (1D size) Motion

Tilt/angle Shape

Area (2D size)

Depth (3D position)

Color luminance

Color saturation

Curvature

Volume (3D size) Relative vs absolute judgement

• Weber’s law says that everything is relative, i.e. the “intensity” depends on the background signal

Unframed Framed

https://thenextweb.com/dd/2015/05/15/7-most-common-data-visualization-mistakes/#.tnw_LZw9olbK © Munzner/Möller 81 http://de.wikipedia.org/wiki/Optische_Täuschung 83 In conclusion

• Visualizations are broken down into marks (elements) and channels (parameters) • Data is linked to these channels • Alot of care in choosing which and how many channels to use

84 Encode: Single View Methods

• Munzner, “Visualization Analysis and Design”: – Chapter 7 (Encode: Arrange in Space) – Chapter 8 (Encode: Spatial Layouts and Link Marks) – Chapter 9 (Encode: Color and Other Channels)

• heavy focus on spatial position for visual encoding • long history for paper-based views of data, see http://www.datavis.ca/milestones/ • many ways to make interactive • many ways to reﬁne / improve / combine

• categorical: bar charts – aligned / ordered

• quantitative/ ordered: dot plot / line chart

• invented by William Playfair (1759-1823) – also invented bar charts, pie charts, etc.

http://www.math.yorku.ca/SCS/Gallery/images/playfair-wheat1.gif

• line implies trend, not appropriate for categorical data

Zacks and Tversky. Bars and Lines: A Study of Graphic Communication. Memory and Cognition 27(6):1073-1079, 1999. © Munzner/Möller 90 2D keys - scatterplots

• encode two input variables with spatial position – show positive/ negative/ no correlation between variables – show clusters: clumpiness/density,

shape, overlap http://upload.wikimedia.org/wikipedia/commons/0/0f/Oldfaithful3.png

© Munzner/Möller 91 Scagnostics e on scerya ule nms ftepnl.Tescatterplot The panels. the The of most uncover. in can ex- outlier scagnostics an an clearly that give is anomalies to point of red SPLOM type graph-theoretic the have the of We ample in the [43]. point by Nations a compiled United highlighted the countries and selected Organization for Health World statistics on based dataset clumpy, lengths, edge in stringy. scatter- skewed and this outliers, striated, characterize in nonconvex, can high we relatively SPLOM, ﬁgure. as the the of plot in right see upper we the what in unusual shown From one is represents plot that linked We SPLOM The dataset. the scatterplot. Abalone in the point for a highlighted SPLOM have scagnostics a shows 5 Figure need the compute. are of to distributions because faster nonuniform considerably compact, time with computation Datasets increase rebinning. for to tend sets point for red=high) patterns (blue=low, scatter measures eleven graph-theoretic Scaled 3: Figure 5E iue6sosasansisSLMo 7vralsfo a from variables 17 of SPLOM scagnostics a shows 6 Figure XAMPLES

L. Wilkinson, A. Anand and R. Grossman, "Graph-theoretic scagnostics," IEEE Symposium on Information ltmtie nodrt aete oeacsil olnigand lensing to accessible more scatter- them sort make to commu- to methods order these learning use in manifold can matrices This we plot the Second, [6]. by [4], space. [32], taken gen- nity high-dimensional those more in follows the problem to and approach detection them these pattern from apply models and eral multivariate how- measures do, construct detecting graph-theoretic to can of other more we is possibility First, There the matrices. ever. offers scatterplot scagnostics large in seen, anomalies have we As distributions. the clusters 2D sorting on and marginal The variables similar loadings continuous and the measures. discrete scagnostic of the sorted segregates the size clearly have of the We component using principal SPLOM first agencies. the Network in federal Climate variables these Greenland the by the sponsored of NSF. part (GC-Net) and are NASA measurements by Greenland operated These the station at weather year automatic a Humboldt over measurements meteorological hourly focus method. screening appropriate preliminary the a illlustrates as application scagnostics of This mea- analysis). to the due in be ( might artifacts by clusters findings surement that authors’ is possibility the the scatters the to Instead, support downweighting 2D dimensions. some the higher lends in in SPLOM clusters scagnostic clusters must of We lack for for evidence data. evidence these of not clus- in lack distribu- identified cells that 2D authors cancerous remember typical The and non-cancerous a of normal. shows lines ters bivariate scatterplot of cell not linked degree 62 evidently The high the tion, of the array. distributions shows the marginal SPLOM in 2D The the in [2]. homogeneity dataset microarray have mining. might data and automated analysis thoughtless statistical a not in a have undetected in been would singularity they a correlated, induced perfectly artifactually- have one two not – these are government Because variables versions of related categories. two type fewer included have fig- the had to the measuring we recoded of variable that right categorical find upper a to the of surprised in were shown We is ure. point this to corresponding

6C Visualization, 2005. INFOVIS 2005., 2005, pp. 157-164.

iue8sosaSLMo ete aa h aacomprise data The data. weather of SPLOM a shows 8 Figure doi: 10.1109/INFVIS.2005.1532142 a from variables 62 of SPLOM scagnostics a shows 7 Figure ONCLUSION © Munzner/Möller 92 iue4 ilto esrsadscatters and measures of Biplot 4: Figure e.g. iigdsrt n otnosvariables continuous and discrete mixing , multiple keys - partitioning / subdivide • e.g. 2 keys – use two perpendicular axis OR – use alignment on one axis • separate by A ﬁrst and then by B (left) • separate by B ﬁrst and then by A (right) • also known as dimensional stacking

• Spatial channel – quantitative vs. categorical attributes – Keys: the importance of ordering • list (1D) vs. matrix (2D) vs. partition / subdivide (multiple D) – Spatial layout • rectilinear • parallel • radial – Spaceﬁlling – Dense • Linemarks – Connection – Containment

• rectilinear: standard • parallel: parallel coordinates • radial: starplots, etc.

• one scatter plot: choose two dimensions • SPloM: Scatter Plot Matrix – show all combinations

• raw, ﬁlter

• only 2 orthogonal axes in the plane • instead, use parallel axes!

• geometric interpretations – hyper plane, hypersphere – points do have intrinsic order • nominal / categorical data – no intrinsic order, what to do? – indeterminate/arbitrary order • weakness of many techniques • downside: human-powered search • upside: powerful interaction technique • most implementations – user can interactively swap axes © Munzner/Möller 99 PC: Axis Ordering

• PC – shows only a subset of axis combinations – relatively compact – has a learning curve • SPloMs – show all combinations of 2 attributes – demanding on screen real-estate – intuitive / known to a broad audience

• use angular channel for dimensions • rectilinear bar chart vs. radial star plot vs. radial multipodes plot

• relevant for drawing trees • containment -- essentially treemaps • connection -- traditional view of graphs • type of information is different

• avoid hairball effect

information density: 10x better H3 PARC Tree

center fringe

3D dozens thousands