Data visualisation in practice

Data visualisation in practice is an optional module. It shares the year three elective slot with HCI3107 Professional practice and the HCI profession. A student takes exactly one of the two, and both carry the same 15 credits. The difference is the subject, not the weight. HCI3107 is about how design work is scoped, sold and delivered. This module is about making visual evidence do its job for an audience that did not make it.

The distinction that governs everything below is between encoding data correctly and making a chart work for someone else. Encoding is a solved problem in principle: the perceptual ranking of channels, the rules for colour and scale, and the catalogue of distortions are all in the literature and are taught at master's level in HCI4104. What is not solved is the second half. A figure is commissioned by one person, drawn by another, read by a third who has ten seconds and no context, and then used to decide something that affects people who never saw it. This module teaches the second half.

A chart answers a question, and the question is usually someone else's

Every visualisation encodes an answer, but the answer is only useful if it addresses a question a reader actually holds. Munzner's framework begins with three questions in a fixed order: what data does the reader need to see, why do they need it, and only then how should it be drawn (Munzner, 2014). The nested model that sits above it exists because errors at the wrong level are expensive. In the nested model, a design that solves a problem the reader never had cannot be rescued by better colour or a cleaner layout, because it fails at the task level, and the later levels depend on the earlier ones (Munzner, 2014).

The commissioning conversation is where this failure is made or avoided. The person asking for a chart usually brings a request rather than a question. They want a dashboard, or a map, or the same figure the board saw last quarter. The designer's job is to convert the request into a named judgement: a comparison, a ranking, a trend, a distribution or an outlier. The work of the unit is to practise that conversion on real briefs, because a chart that answers a question nobody asked is not a lesser chart, it is a different object.

The medium constrains the form before the data does

A form that works on a laptop fails on a phone, a slide, and a printed page. The reason is perceptual rather than aesthetic. The ranking of channels by the accuracy of the judgements they support was established in controlled experiments, and position on a common scale supports the most accurate quantitative judgement while area, angle and colour saturation support the least (Cleveland and McGill, 1984). Twenty-six years later the same comparisons were run with crowdworkers on the web, and the order survived while two things changed: the most accurate tasks moved closer together, and rectangular area proved surprisingly poor at an aspect ratio of one (Heer and Bostock, 2010).

Those findings decide what survives a change of medium. Shrink a chart to a phone and the small differences that position made legible disappear; a reader can no longer resolve the gaps between bars that carried the argument. Project it on a wall and the opposite happens: text becomes unreadable while gross position and colour survive. Print it in greyscale on cheap paper and any encoding that leaned on hue stops working. This unit takes one dataset and one question and rebuilds the chart for five media, judging each version against the channel ranking rather than against taste.

Colour is an accessibility and reproduction requirement, not a finish

Colour does two different jobs that are often confused. It can carry identity, when categories have no order, and it can carry magnitude, when values are ordered. A palette that is perceptually uniform, meaning equal steps in the data produce equal steps in what the eye sees, is required for the second job, and the first job has an entirely separate set of constraints. The common defaults fail both tests more often than designers expect. Crameri and colleagues catalogued the palette literature and reported that colour maps which distort data through uneven gradients, or which are unreadable for readers with colour-vision deficiency, remained prevalent in science, and they proposed derived maps that report true variation while staying readable (Crameri, Shephard and Heron, 2020).

The accessibility half is a reproduction requirement because it is a legal and professional one, not a preference. Roughly one in twelve men has a form of colour-vision deficiency, and the number of readers affected by greyscale printing, low-quality projectors and bright screens is larger still. This unit therefore asks for every palette to be tested under simulation and in greyscale before it is used, and for a chart to keep working when the colour is removed entirely.

Uncertainty is where published charts fail most often

An estimate drawn without its interval invites a reader to treat it as a fact. Drawing the interval is necessary but not sufficient, because the representation changes what the reader infers. Hullman and colleagues surveyed the field and set out a taxonomy of the decisions an evaluation of an uncertainty display must make, from the behavioural target down to the analysis, and then applied it to eighty-six user studies (Hullman, Qiao, Correll, Kale and Kay, 2019). They found that evaluation practice concentrates on performance and satisfaction measures, and that those measures assume more predictable judgement than the research on judgement and decision making supports.

The practical consequence is uncomfortable. A display that shows uncertainty can score worse on a standard abstract task and still improve a decision, and a clean interval that is statistically correct can make a reader more confident about the direction of an effect than the data warrants. This unit teaches the range of established remedies, from error bars and violin plots to hypothetical outcome plots, and it asks you to say which inference each one protects and which it exposes.

Annotation and sequence are how a static figure carries an argument

A chart that only shows data leaves the reading to the reader. A chart that carries an argument selects, annotates and orders, and the research on narrative visualisation shows how much of the author's reading enters at that point. Segel and Heer reviewed fifty-eight published examples and derived seven genres, among them the magazine style, the annotated chart and the slide show, organised by the balance between an author-driven narrative and reader-driven exploration (Segel and Heer, 2010). Hullman later showed that the same accurate data can support opposite conclusions depending on how the sequence is framed, which is why this module treats annotation as a rhetorical act and asks you to disclose the reading you have built in.

Maps, rates and small numbers manufacture patterns

A map is not a neutral container. It encodes geography alongside data, and geography brings three traps that recur in professional work. The first is the choice of areal unit: the same underlying values produce different correlations when they are aggregated into different boundaries, which is the modifiable areal unit problem. The second is the ecological fallacy, the error of reading a relationship between areas as a relationship between the people who live in them, a mistake named by Robinson seventy years ago and still made in dashboards today (Robinson, 1950). The third is the instability of small counts, where a rate computed from a handful of cases can top a ranking and then fall away in the next period.

The remedies are established. Rates are preferred to counts for comparison, counts are shown rather than hidden, unstable rates are smoothed or shown with their interval, and the choice of unit is stated. This unit asks you to build the same dataset as a count map and a rate map and to explain, with the numbers, why a reader could reach two different conclusions.

Health and risk communication is where the design claims are tested

The evidence base for visual communication is thin almost everywhere except one area: health and risk communication, where randomised studies have tested what readers actually understand. That literature is unusually specific. Gigerenzer and colleagues reviewed the evidence on how doctors and patients make sense of health statistics and found widespread difficulty with conditional probabilities, and they showed that the same information presented as natural frequencies rather than percentages, and as absolute rather than relative risk, is understood by markedly more people (Gigerenzer, Gaissmaier, Kurz-Milcke, Schwartz and Woloshin, 2007). Ancker and colleagues reviewed the design features of graphs used in health risk communication and found that although graphics have a strong theoretical basis, the accumulating evidence is mixed, and that design features such as use of colour, scales and the presence of a reference point can change comprehension and preference in ways that editors rarely control (Ancker, Senathirajah, Kukafka and Starren, 2006).

This unit uses that literature as the strictest available test of the design claims in the rest of the module. If a frequency format reliably beats a percentage format in a trial, then the general claim that format matters has at least one place where it was measured rather than asserted.

Dashboards are instruments inside organisations, judged by decisions

A dashboard is not a large chart. It is an instrument with an owner, a cadence and a decision attached, and that is why its failures are organisational as well as visual. Sarikaya and colleagues asked what people actually talk about when they talk about dashboards, and found a body of practice held together by convention rather than definition, with design considerations that are rarely made explicit (Sarikaya, Correll, Bartram, Tory and Fisher, 2019). A panel that is measured changes behaviour, and a measured panel invites the organisation to improve the number rather than the thing it was meant to represent. This unit connects dashboard design to the performance-measurement literature, and it asks you to cost a wrong panel: which decision it would distort, who would feel it, and how the distortion would be detected.

Evaluating a visualisation means choosing what counts

Evaluation forces the question that everything above circles. Comprehension, trust, decision quality and insight are different targets, and a method that measures one says little about the others. An insight-based method counts the observations a reader makes in a think-aloud protocol rather than scoring answers to a set question, which measures something closer to what analysts do, at the cost of being labour intensive and hard to standardise (Saraiya, North and Duca, 2005). A comprehension test is repeatable but can only report on the search it taught. A trust measure captures confidence, which is not the same as appropriate reliance.

This is also where the ethics of the module sit. A display can be accurate and still lead a reader to a conclusion the author wanted, and the difference between informing and persuading is a design decision, not an accident. Correll argues that visualisation research carries ethical responsibilities that its methods rarely examine, from the framing of a result to the consequences of a display that will be read as fact (Correll, 2019). The module asks you to state what your evaluation measured, what it could not measure, and what a reader might reasonably take from your chart that you did not intend.

The next stage is the ten units themselves. Each one takes a slice of this literature, argues a position, names where the evidence disagrees, and ends with questions that test whether you can predict what a given design will do to a given reader.