Data Visualisation & Perception

Visualisation is often treated as the pretty step at the end — make a chart, tidy the colours, ship it. That undersells it badly. A well-made chart lets a reader see a pattern, a gap, or an outlier instantly that would take paragraphs to describe and a table to hide. It works because the human visual system is a massively parallel pattern-finder, and a good chart hands the work to it. A bad chart fights it.

So the right way to think about visualisation isn't aesthetics — it's perception. Once you know what the eye does automatically, what it judges accurately, and what it gets wrong, the rules for good charts stop being arbitrary taste and become something close to engineering. That's the through-line of this page.

Why we draw data

The eye and visual cortex process enormous amounts of information in parallel and pre-consciously — far more than the slow, serial channel we use to read numbers. A table of a thousand rows is a thousand serial reads; the same data as a scatter plot is one glance. Visualisation converts a cognitive task (compute, compare, remember) into a perceptual one (look, see), and perception is the faster, higher-bandwidth system.

The famous demonstration is Anscombe's quartet: four datasets with nearly identical means, variances, and correlation, that look completely different when plotted — one linear, one curved, one a single outlier dragging a line. The summary statistics hide what one chart reveals. That's the case for visualisation in a nutshell: the picture carries structure the numbers flatten away.

Pre-attentive processing: what the eye does for free

Some visual properties are processed pre-attentively — automatically, in a fraction of a second, before conscious attention engages. A single red dot in a field of grey ones "pops out"; you don't search for it, you just see it. These pre-attentive attributes include colour (hue), size, orientation, position, and shape.

This is the most powerful lever in visualisation, because it's effectively free attention. Encode the thing you want noticed in a pre-attentive attribute and the reader notices it instantly. The flip side is the warning: if everything is bold and colourful, nothing pops — you've spent the budget on noise. Pre-attentive emphasis only works when it's scarce.

The accuracy ladder

Not all ways of encoding a number are equally readable. Cleveland and McGill ran the experiments and ranked visual encodings by how accurately people judge the underlying quantity. The order, most accurate first:

The Cleveland-McGill ranking. We read position on a common scale most accurately, then length, then angle and slope, then area, then colour and density. Match the most important comparison to the highest-accuracy encoding you can.

The practical rule falls straight out: match your most important comparison to the highest-accuracy encoding available. This is the real reason a bar chart usually beats a pie chart — comparing bar lengths/positions is easy and precise, comparing pie angles is hard and error-prone. It's not snobbery; it's that one asks more of the reader's eye than the other.

Choosing the chart: start from the question

The chart type isn't a style choice — it follows from what question you're answering. A quick map:

Comparison (which is bigger?) → bar chart, dot plot.
Trend over time → line chart.
Relationship (do two variables move together?) → scatter plot.
Distribution (what's the spread?) → histogram, box plot, density.
Composition (parts of a whole) → stacked bar; a pie only for a couple of slices, and even then reluctantly.

Decide the question first, pick the encoding that reads most accurately for it, and only then worry about looks. A beautiful chart answering the wrong question is still the wrong chart.

Colour, done right

Colour is powerful and easy to misuse. The first rule is to match the type of colour scale to the type of data:

Sequential — light-to-dark of one hue, for ordered/quantitative data (low to high). More is darker.
Diverging — two hues meeting at a neutral midpoint, for data with a meaningful centre (above/below zero, above/below target).
Categorical — distinct hues for unordered groups; keep it to a handful, since people can't track many colours at once.

Chart crimes & honesty

Because charts are so persuasive, they're easy to use to mislead — sometimes on purpose, often by accident. The common offences:

Truncated axes. Starting a bar chart's y-axis above zero exaggerates small differences into dramatic ones. Bars encode length, so they must start at zero. (Line charts of an index have more latitude, but label it clearly.)
Dual y-axes. Two different scales on one chart lets you manufacture a "correlation" by sliding the axes until the lines align. Usually best avoided.
3D and decoration. 3D pie charts and perspective distort the very areas/angles they encode. The decoration actively corrupts the data.
Overplotting. Thousands of points piled into an opaque blob hides the density. Use transparency, binning, or sampling so the structure shows.

The honest test: would a reader glancing for two seconds come away with the true takeaway? If the visual encoding pushes them toward a wrong conclusion, the chart is lying even if every number is correct.

Data-ink & decluttering

Edward Tufte's enduring idea is the data-ink ratio: of all the ink (pixels) on a chart, what fraction actually encodes data versus decoration? Maximise it. Every gridline, heavy border, background fill, drop shadow, and redundant label is chart junk competing with the signal for the reader's attention.

Decluttering is mostly subtraction: mute or remove gridlines, drop the chart border, label directly instead of via a distant legend, and delete anything that doesn't help the reader answer the question. The goal is the opposite of "more impressive" — it's that the data is the loudest thing on the page. This connects straight to communicating clearly: a decluttered chart is a clear sentence, not a busy paragraph.

Where it shows up in my work

Refresh in 60 seconds

A chart borrows the eye's bandwidth — it turns a cognitive task into a perceptual one. Anscombe's quartet: plot it, don't just summarise it.
Pre-attentive attributes (colour, size, position) are noticed instantly — powerful, but only if used scarcely.
Accuracy ladder(Cleveland-McGill): position > length > angle > area > colour. Match the key comparison to the highest encoding — why bars beat pies.
Start from the question: comparison→bar, trend→line, relationship→scatter, distribution→histogram/box, composition→stacked bar.
Colour: sequential / diverging / categorical to match the data; never red-green only; design colour-blind-safe. Don't encode precise quantities in colour.
Honesty: bars start at zero, avoid dual axes & 3D, fix overplotting. Maximise data-ink, cut chart junk.

The encoding ranking (Cleveland-McGill), pre-attentive processing, and colour-accessibility guidance reflect current data-visualisation and perception references alongside coursework.