Data visualisation and visual analytics

A chart is an instrument for a judgement. It exists so that a reader can compare, rank, estimate or notice something, and the encoding decides which judgements are easy and which are nearly impossible. It is also where the field argues with itself. The perceptual evidence comes from abstract judgement trials, the craft rules from designers reading their own examples, and the ethical claims from journalists.

A mark and a channel, and why the ranking is about judgement

Data has two parts: items, the things described, and attributes, what is known about each. A mark is the geometric thing you draw: a point, a line, an area or an icon. A channel is the visual property that controls it, such as position, length, area, angle, slope, colour or motion.

Munzner built the framework the field now teaches from (Munzner, 2014). Design reduces to three questions. What data does the reader need to see, why do they need it, and how can marks and channels carry it. That order explains why a chart cannot be chosen before its question is known.

The channel ranking is the framework's best-known part. Position on a common scale supports the most accurate quantitative judgement, length on a common scale comes next, and area, angle and colour saturation support the least accurate. The evidence is experiments in which readers saw two marks and estimated what percentage the smaller was of the larger (Cleveland and McGill, 1984). From the errors the authors derived elementary perceptual tasks, ordered them by accuracy, and drew one guideline: use the highest tasks you can. Their paper is careful about its own limit, noting that the experiments also showed the set of tasks had to be expanded.

The ranking survives a second test, with exceptions

Twenty-six years later, Heer and Bostock ran the same comparisons with crowdworkers on the web, using the same error measure so the results could sit side by side (Heer and Bostock, 2010). The shape and order of the judgement types were preserved. Two things changed. The most accurate judgement types came closer together, and length judgements improved on the original study. Position still significantly outperformed length.

They also went further. In a new experiment on rectangular area, of the kind treemaps and cartograms produce, accuracy matched that of circular area. Aspect ratio mattered significantly, and the surprise was the direction. Pairs with an aspect ratio of one, the squares a squarified treemap aims for, produced the worst judgements. Readers appear to use one-dimensional length as a proxy for area, and the sides of two squares give that proxy nothing to work with.

The limits matter here. Display, lighting and viewing distance were uncontrolled, so only a subclass of perception experiments can run this way. The authors add that visual variables interact, so extrapolating from isolated variables is unreliable. That is the contested question stated plainly. A ranking built from abstract judgements predicts the direction in which a real chart fails, and it does not tell you what to draw.

You cannot choose an encoding before you know the question

Task abstraction is the discipline of naming the judgement before drawing anything. A reader asking whether one value exceeds another is making a comparison. A reader asking how something changed is reading a trend. A reader asking whether values cluster is reading a distribution.

Take twelve months of waiting times at one emergency department, a value per month. As a comparison of two months, a pair of bars on a common baseline is read by length. As a trend, the same values are read by slope in a line chart. As a distribution, the interest moves to how many months fell above the target, which a dot plot answers. As an outlier hunt, the months belong on a sorted dot plot, where position on a common scale does the work. One dataset, four charts, each defensible only because a judgement has been named.

The test is cheap. If you cannot say which judgement a chart supports, it has no defence beyond taste, and decoration accumulates where an argument belongs.

Colour is three problems wearing one name

Colour is categorical when the values have no order, sequential when they are ordered and run one way, and diverging when they are ordered and meaningful in two directions from a centre. Hue carries identity rather than magnitude, so a categorical palette used for a count destroys the ordering the reader needs.

A sequential ramp needs perceptual uniformity: equal steps in the data must produce equal steps in what the eye sees. Judged in lightness, most naive ramps fail. A rainbow palette climbs in lightness quickly in the middle and flattens at the ends, which manufactures sharp-looking boundaries where the data is smooth.

Crameri, Shephard and Heron catalogue the practice and the cost (Crameri, Shephard and Heron, 2020). Their abstract states that colour maps which distort data through uneven gradients, or which are unreadable for readers with colour-vision deficiency, remain prevalent in science, and names rainbow-like and red-green maps among them. Their remedy is a set of scientifically derived maps that report true variation and stay readable.

The evidence deserves care. That article is an argument and a guide with reference palettes, not a controlled trial of reader accuracy. The colour-vision claim rests on the settled physiology of the cone pigments plus simulation, which is a model. The distortion claim rests on each palette's measurable lightness profile.

Axes, baselines and the arithmetic of distortion

Most misleading charts are defaults that were never examined, and the mechanism is usually arithmetic. A bar encodes magnitude as length, and length is proportional to magnitude only when the baseline is zero.

Truncate a bar chart axis at eighty per cent and a bar for eighty-two units is drawn as one tenth the height of a bar for a hundred. A line chart is different, because slope carries the reading, so truncation exaggerates apparent change without destroying any drawn length. A second pair of axes, scaled independently, can manufacture a relationship out of two free choices of scale. Area encodings fail similarly, since doubling the radius of a circle quadruples its area.

Cairo argues that truthfulness is a continuum rather than a property (Cairo, 2016). His book separates the claim into stages, through the truth continuum and conjecture and uncertainty to confidence and significance, and it offers a step-by-step process for evaluating a graphic.

Tufte made the purist version of the argument (Tufte, 1983). His book sets out the data-ink ratio, treating as waste any ink that does not carry data, argues for small multiples, and proposes detecting graphical deception by comparing design variation with data variation. Read critically, the ratio is a heuristic, illustrated with examples chosen by its author rather than tested. Erasing gridlines and tick labels can lower the accuracy of the estimate the reader has to make, which the perceptual ranking protects. The argument survives as a discipline of attention, not as a law.

Interaction is how a view is made manageable

Shneiderman's paper supplied the phrase the field still uses: overview first, zoom and filter, then details on demand (Shneiderman, 1996). Its abstract says this is a starting point, and it offers seven data types and seven tasks: overview, zoom, filter, details on demand, relate, history and extract.

Consider four hundred thousand patient records and a question about one subgroup. The overview shows the shape of the whole, usually as a set of distributions in a small multiple. Selection filters, and every other view updates to the same subset, which is what brushing and linking means. Details on demand brings one record into reach, and extract lets the finding leave the tool.

Small multiples are the quiet workhorse. The same encoding repeats in a grid with fixed scales, so a reader compares panels by eye and reads values within a panel. That separates two judgements instead of asking one plot to carry both.

Coordinated views introduce a risk that is easy to miss. Once several views filter together, a relationship visible in one panel can be an artefact of the selection, and the analyst has no way to see the selection as part of the data.

Dashboards fail as documents and work as instruments

A dashboard is a genre with an audience, a cadence and a decision, not a shelf of charts. The failure modes are recognisable. It answers no named question. It gives twelve panels equal weight, so nothing is salient. It hides the definition of the metric, and a readmission rate can be defined three ways. It reports a value with no reference band, so the reader cannot tell whether it is good.

Take an operational dashboard for a hospital. A first version shows a grid of sixteen measures, each with a small gauge. A workable version keeps the data and changes the reading. A short row of headline measures carries an explicit expected range, drawn as a band, so each number is compared with something. Every headline carries a sparkline over the previous ninety days, which supplies the trend. The ward comparison moves to one common scale and one sorted order, so ranking is read by position.

None of that is decoration: every change serves a judgement a named reader has to make. That guidance rests on the framework and on the perceptual experiments, not on a body of dashboard studies.

Uncertainty is where published charts fail most often

An estimate drawn without its interval invites a reader to treat it as a fact. An interval drawn as a bar or a band uses position and length, the channels that support accurate comparison. A distribution drawn in full, rather than summarised by a mean, shows the shape a summary hides. Frequency framing, in which a count of icons stands for a count of cases, lets a reader count rather than judge an area.

Hullman and colleagues surveyed how far the field has got in testing those remedies (Hullman, Qiao, Correll, Kale and Kay, 2019). They set out a taxonomy of six levels of decisions an evaluation of an uncertainty display has to make, from the behavioural target to the analysis. Applied to eighty-six user studies, the survey found that evaluation practice, particularly in visualisation research, concentrates on performance and satisfaction measures. Those measures assume more predictable judgement than the research on human judgement and decision making supports. The authors also report a bias toward reading performance as accuracy rather than as decision quality.

That finding is why this section is hard. A display that shows uncertainty can score worse on an abstract task and still improve a decision, and a standard evaluation will not see the difference.

Narrative, annotation and the difficulty of measuring comprehension

Narrative visualisation adds an author's sequence to a reader's exploration. Segel and Heer reviewed fifty-eight published examples and derived seven genres, among them the magazine style, the annotated chart, the slide show and the video (Segel and Heer, 2010). Their organising idea is a balance between author-driven narrative flow and reader-driven story discovery. Annotation is how the author's reading enters a display, and it is also where persuasion enters.

Evaluation decides whether any of it worked, and this is where the literature is weakest. A controlled task measures accuracy and time on a judgement the experimenter chose, so the method reports only on the search it taught. Saraiya, North and Duca attacked the problem directly (Saraiya, North and Duca, 2005). They defined an insight as an individual observation about the data and counted insights in a think-aloud protocol, rather than scoring answers to a set question. Thirty participants each used one of five microarray tools on one of three datasets.

Measured insight did not depend on the participant's background, so domain experts generated no more insight than novices. Believed insight did depend on background: software developers felt they had learned less, and novices felt they had learned more. And multi-view tools, which linked several views to a selection, gave users confidence they had missed nothing.

The authors also wrote down the costs. Coding insight is labour intensive, it needs a domain expert coder, it needs self-motivated participants, and each session lasted under two hours. A person analysing their own data works for days, weeks or months, and the authors say long-term insight may differ. Their recommendation is a study that follows analysts through repeated cycles on their own data, which is where visualisation becomes a question about use over time.