Critical Design in Statistical Visualization Alan Blackwell Professor of Interdisciplinary Design University of Cambridge

Overview Brief background Principles of visualisation design Four critical case studies The Automated Statistician ICUMAP Gatherminer Self-Raising Data

Background: Metaphor in Diagrams

Some Principles of visualization

Typography and text

Maps and graphs

Schematic drawings

Node-and-link diagrams

Icons and symbols

Visual metaphor

Microsoft “Bob” (1995)

Microsoft “Task Gallery” (2000)

MIUI “Warm Space MiHome Desktop” (2015)

Pictures

Unified theories of visual representation

Graphic Resources Correspondence Design Uses Marks Shape Orientation Size Texture Saturation Colour Line Literal (visual imitation of physical features) Mapping (quantity, relative scale) Conventional (arbitrary) Mark position, identify category (shape, texture colour) Indicate direction (orientation, line) Express magnitude (saturation, size, length) Simple symbols and colour codes Symbols Geometric elements Letter forms Logos and icons Picture elements Connective elements Topological (linking) Depictive (pictorial conventions) Figurative (metonym, visual puns) Connotative (professional and cultural association) Acquired (specialist literacies) Texts and symbolic calculi Diagram elements Branding Visual rhetoric Definition of regions Regions Alignment grids Borders and frames Area fills White space Gestalt integration Containment Separation Framing (composition, photography) Layering Identifying shared membership Segregating or nesting multiple surface conventions in panels Accommodating labels, captions or legends Surfaces The plane Material object on which the marks are imposed (paper, stone) Mounting, orientation and display context Display medium Literal (map) Euclidean (scale and angle) Metrical (quantitative axes) Juxtaposed or ordered (regions, catalogues) Image-schematic Embodied/situated Typographic layouts Graphs and charts Relational diagrams Visual interfaces Secondary notations Signs and displays

Analysis examples

Graphic Resources Correspondence Design Uses Marks Size Colour Mapping (quantity, relative scale) Mark position identify category (colour) Express magnitude (size) Symbols Geometric elements Connective elements Topological (linking) Diagram elements Visual rhetoric Regions Alignment grids Containment Separation Framing (composition) Segregating or nesting multiple surface conventions in panels Accommodating labels, captions or legends Surfaces Display medium (web browser) Metrical (quantitative axes) Image-schematic? Graphs and charts

Graphic Resources Correspondence Design Uses Marks Shape Conventional (arbitrary) Mark position identify category (shape) Symbols Geometric elements Letter forms Connective elements Topological (linking) Acquired (specialist literacies) Texts Definition of regions Regions Alignment grids White space Containment Separation Segregating and nesting multiple surface conventions in panels Accommodating labels Surfaces Material object on which the marks are imposed (paper) Metrical (quantitative axes) Juxtaposed and ordered (regions) Musical score

“Big data” = too much data to see

Automated statistician https://www.automaticstatistician.com/static/assets/auto-report-affairs.pdf

Automated statistician output 3. Model description In this section I have described the model I have constructed to explain the data. A quick summary is below, followed by quantification of the model with accompanying plots of model fit and residuals. 3.1. Summary. The output affairs decreases linearly with input religiousness increases linearly with input yearsmarried decreases linearly with input age increases linearly with input occupation decreases linearly with input education

Automated statistician output Decrease with religiousness The correlation between the data and the input religiousness is -0.22 (see figure 2a). This correlation does not change when accounting for the rest of the model (see figure 2b). Increase with yearsmarried The correlation between the data and the input yearsmarried is 0.16 (see figure 3a). Accounting for the rest of the model, this changes slightly to a part correlation of 0.24 (see figure 3b).

Automated statistician output Low negative deviation between quantiles of test residuals and model There is an unexpectedly low negative deviation from equality between the quantiles of the residuals and the assumed noise model (see figure 8a). The minimum of this deviation occurs at the 99th percentile indicating that the test residuals have unexpectedly light positive tails. The minimum value of this deviation is -6.7 which is moderately lower than its median value under the proposed model of -1.9 (see figure 8c). To demonstrate the discrepancy in simpler terms I have plotted histograms of test set residuals and those expected under the model in figure 8b.

Automated statistician – critical notes Inspired by Turing Test (and Lenat’s AM) Actual report text is clearly written by a human Key ability is scanning for interest (what is interesting on average? blue?) Repair, Attribution and all That The reader scans for interest Names of the variables are essential How to pass the Turing Test? Lower the threshold! Reduce expectations? Or offer mixed initiative

Mixed initiative visualization

ICUMAP http://www.cl.cam.ac.uk/~as2388/patients/web_content/

ICUMAP – critical notes Big data measures too many things to visualise Provide ‘dimensionality reduction’ From ‘clusters’ to abstract ‘landscape’ Journey metaphor Use colour classes for a gestalt view Encourage guided browsing Support human expert judgment Show values within distributions Reveal broad spread of ‘similarity’ Use overlays to compare original variables

Gatherminer http://advaitsarkar.github.io/Gatherminer/

Gatherminer – critical notes Use visual rhythm to think about time Mundane comparison is automated Collect for pattern A human is the judge of interest Diagnosis through density But related to individual cases

Self-Raising Data http://www.cl.cam.ac.uk/~mcm79/srd/self-raising-data.html

Self-Raising Data – critical notes How can we do big data without data? Real data is produced by questions Offer data as imagination Rehearse statistical reasoning Structured in a way statisticians can understand subverts the ‘automated statistician’ fantasy Are statistics objective? AI is always (normatively) subjective Uncertainty in information is seen as noise

Summary Visualisation is a design discipline The marks we make communicate and persuade When statistics becomes artificial intelligence … What are we trying to prove? Where does the judgment take place? Mixed initiative interaction Respects expertise Revealing, not concealing, the models used