VISUAL ENCODING & DESIGN PROCESS Miriah Meyer School of - PowerPoint PPT Presentation

VISUAL ENCODING & DESIGN PROCESS Miriah Meyer School of Computing University of Utah

2

Visualization Design Lab @ the vdl.sci.utah.edu

- visual encodings - guidelines - process

visual encodings

how can you visually represent the numbers 4 and 8?

where do these rankings come from?

al field that comes without apparent mental effort. We also perform cognitive tasks such as reading scale information, but much of the power of graphs-and what distinguishes them from tables- comes from the ability of our preatten- Graphical Perception and Graphical tive visual system to detect geometric patterns and assess magnitudes. We for Analyzing Scientific Data Methods have examined preattentive processes Crowdsourcing Graphical Perception: Using Mechanical rather than cognition. We have studied the elementary Turk to Assess Visualization Design William S. Cleveland and Robert McGill tasks theoretically, graphical-perception ideas from the more general borrowing field of visual perception (7, 8), and Jeffrey Heer and Michael Bostock experimentally by having subjects judge Computer Science Department elements (1, 5). The next two Graphs provide powerful tools both mation from graphs; theory and experi- graphical Stanford University data and for com- mental data are then used to order the sections illustrate the methodology for analyzing scientific with { jheer, mbostock } @cs.stanford.edu a few examples. municating quantitative information. tasks on the basis of accuracy. The or- data The computer graphics revolution, has an dering important application: which began in the 1960's and has inten- should be encoded so that the visual the past several years, stim- decoding involves tasks as high in the sified Study of Graphical Perception: Theory during ulated the invention of graphical meth- as possible, that is, tasks per- ordering for ecological validity. Crowdsourced experiments may also ABSTRACT Figure 2 provides an illustration of substantially reduce both the cost and time to result. Understanding perception is critical to effective visualiza- theoretical that borrows some reasoning tion design. With its low cost and scalability, crowdsourcing Unfortunately, crowdsourcing introduces new concerns to be perception is the visual decoding of the quantitative ideas from the field of computational and Summary. Graphical presents an attractive option for evaluating the large design addressed before it is credible. Some concerns, such as eco- information encoded on graphs. Recent investigations have uncovered vision (8). Suppose that the goal is to qualitative space of visualizations; however, it first requires validation. basic principles of human that have important judge the ratio, r, of the slope of line logical validity, subject motivation and expertise, apply to for graphical perception implications In this paper, we assess the viability of Amazon’s Mechanical the display of data. The computer revolution has stimulated the invention BC to the slope of line segment any study and have been previously investigated [13, 14, 23]; of graphics segment Turk as a platform for graphical perception experiments. We methods for analyzing and presenting AB in each of the three panels. Our scientific data, such as box others, such as display configuration and viewing environ- many graphical replicate previous studies of spatial encoding and luminance two-tiered error dot charts, and graphing on a log visual system tells us that r is greater ment, are specific to visual perception. Crowdsourced per- bars, scatterplot plots, smoothing, contrast and compare our results. We also conduct new ex- base 2 scale. than 1 in each panel, which is correct. ception experiments lack control over many experimental periments on rectangular area perception (as in treemaps or Our visual system also tells us that r is conditions, including display type and size, lighting, and cartograms) and on chart size and gridline spacing. Our re- closer to 1 in the two rectangular subjects’ viewing distance and angle. This loss of control panels sults demonstrate that crowdsourced perception experiments than in the square panel; that is, the ods: types of graphs and types of quanti- formed with greater accuracy. This is inevitably limits the scope of experiments that reliably can are viable and contribute new insights for visualization de- tative information to be shown on graphs illustrated by several in which slope of BC appears closer to the slope examples be run. However, there likely remains a substantial subclass sign. Lastly, we report cost and performance data from our (1-4). One purpose of this article is to of AB in the two rectangular some much-used graphical forms are than panels of perception experiments for which crowdsourcing can pro- experiments and distill recommendations for the design of describe and illustrate several of these presented, set aside, and replaced by in the square panel. This, however, is vide reliable empirical data to inform visualization design. crowdsourced studies. new methods. new methods. r is the same in all three pan- incorrect; In this work, we investigate if crowdsourced experiments in- els. What has been missing, until recently, H5.2 [Information interfaces and pre- ACM Classification: sensitive to environmental context are an adequate tool for in this of rapid The reason for the distortion in invention period judging graphical sentation]: User Interfaces—Evaluation/Methodology Fig. 2 is that our visual system is geared graphical perception research. We assess the feasibility of and deployment is the study of graphs Elementary Tasks for the Graphical Experimentation, Human Factors. General Terms: and the human visual system. When a to judging angle rather than slope. In using Amazon’s Mechanical Turk to evaluate visualizations Perception of Quantitative Information graph is constructed, quantitative and their work on computational theories of and then use these methods to gain new insights into visual- Information visualization, graphical perception, Keywords: categorical information is encoded, The first step is to identify vision in artificial intelligence, Marr ization design. We make three primary contributions: elementary (8) user study, evaluation, Mechanical Turk, crowdsourcing. tasks that are used and Stevens (9) have investigated how chiefly through position, shape, size, graphical-perception • We replicate prior laboratory studies on spatial data en- informa- judge the slant and tilt (10) of the symbols, and color. When a person to visually extract quantitative INTRODUCTION people codings and luminance contrast using crowdsourcing tech- looks at a graph, the information is visu- tion from a graph. (By "quantitative surfaces of three-dimensional objects. we mean numerical values that we judge slant and tilt as ally decoded by the person's visual sys- information" They argue tem. A graphical method is successful of a variable, such as frequency of radia- angles and not, for example, as their gross national that are which are the slopes. An angle only if the decoding is effective. No tion and product, tangents, matter how clever and how technologi- not highly discrete; this excludes cate- contamination of slope judgments ex- such as type of met- the distortion in of Fig. cally impressive the encoding, it fails if information, plains judgments gorical al and nationality, which is also shown 2. Let the angle of a line segment be the the decoding process fails. Informed Ten tasks with which angle between it and a horizontal ray decisions about how to encode data can on many graphs.) we have worked, in our theoretical in- (0 in Fig. 3). The be achieved an understand- to the right only through extending and in our experiments, are of the line segments in the square ing of this visual decoding process, vestigations angles panel of Fig. 2 are not as similar in which we call graphical the following: angle, area, color hue, perception (5). color saturation, density (amount of as the angles in either of the Our second is to convey some magnitude purpose a this makes the slopes recent theoretical and experimental in- black), length (distance), position along rectangular panels; on identical but in the rectangular panels seem closer in of graphical We common scale, positions vestigations perception. nonaligned scales, slope, and volume value. identify certain elementary graphical- Again, let 0 be the angle of a line in perception tasks that are performed (Fig. 1). Visual decoding as we define it for a second line segment the visual decoding of quantitative infor- segment. Suppose has an angle 0 + AO where AO is small tasks is elementary graphical-perception The authors are statistical scientists at AT&T Bell in vision (6): but that a difference what Julesz calls preattentive just large enough 600 Mountain Laboratories, Avenue, Murray Hill, of the line segments can the instantaneous of the visu- the orientations perception New Jersey 07974. VOL. 229 SCIENCE, 828

how much longer is B? A 4x B

how much larger is B? 2x diameter 4x area A B

how much darker is B? 2x A B

1 2

1 2 low time high

high low time 1 2 g1 g2 s6 s7 s1 s5 s3 s4 s2 value

EXERCISE: encoding deconstruction