Analytics and Data Structures using AI

Principles of Data Visualization

A teaching catalogue of rigorous, high-signal visualization principles associated with Edward Tufte, William Cleveland, Alberto Cairo, Stephen Few, the Financial Times graphics desk, and other practitioners. Each principle is paired with online examples that can be discussed in class.

16 principles

What makes a visualization excellent?

A good visualization is evidence arranged for thought. It should make important comparisons visible, preserve the truth of the data, remove distracting ornament, and help the viewer ask better next questions.

Truthful The geometry, scale, and labels do not exaggerate, hide, or distort the data.
Comparative The viewer can compare groups, time periods, places, distributions, and exceptions.
Useful The design answers a real analytic question for a real audience.
1

Begin With a Question

A chart is not a decoration for data. It is a device for answering a question, testing a claim, or provoking a better question.

Use in class

  • Name the decision or idea the viewer should be able to inspect.
  • Choose the chart type after the comparison is clear.
  • Do not ask one chart to answer five unrelated questions.
Ask: after 20 seconds, can a student say what question this chart is trying to answer?

Examples

Gapminder health and wealth bubble chart

Gapminder Health and Wealth

The question is plain: how do income, life expectancy, population, region, and time move together?

purposemultivariate
Open Gapminder Tools
Our World in Data: long-run global change charts

Our World in Data

Charts are built around explainable public questions: poverty, mortality, energy, health, and emissions over time.

public datacontext
Open Our World in Data
Financial Times Visual Vocabulary

FT Visual Vocabulary

A practical catalogue organized by analytic job: deviation, correlation, distribution, ranking, change, and more.

chart choicequestion first
Open the visual vocabulary
2

Show Comparisons, Not Isolated Numbers

Most meaning comes from comparison: against another group, a previous value, a target, an expected value, or a distribution.

Use in class

  • Prefer paired or adjacent views when differences matter.
  • Order categories by value unless natural order is more important.
  • Include reference lines for targets, historical averages, or thresholds.
Ask: what is the natural comparison, and did the chart make that comparison effortless?

Examples

Charles Minard Napoleon campaign map

Minard's Napoleon Campaign

One graphic compares army size, direction, geography, dates, and temperature to explain loss over the campaign.

comparisonmultivariate
Open source
Cleveland dot plots

Cleveland Dot Plot

Dots along a common scale make ranked category comparisons cleaner than heavy bars or slices.

rankingposition
Open explanation
Nightingale mortality diagram

Nightingale Mortality Diagram

Monthly causes of death are compared to make preventable disease impossible to ignore.

public healthcomparison
Open source
3

Preserve Graphical Integrity

The visual effect should be proportional to the numerical effect. Design choices should not create a false story.

Use in class

  • Check whether area, length, angle, or color exaggerates the data.
  • Use baselines and axes that fit the analytic task.
  • Separate genuine signal from noise, missingness, and sampling effects.
Ask: if the chart were evidence in a debate, what design choice could be challenged?

Examples

Challenger O-ring temperature plots

Challenger O-ring Evidence

A classic integrity case: ordering evidence by the causal variable, not by launch date, changes the argument.

causalityethics
Open case
Anscombe quartet

Anscombe's Quartet

Identical summary statistics hide four different patterns. Integrity requires looking at the data, not only summaries.

scatterplotdiagnostics
Open source
Google ML: Visualization traps

Visualization Traps

A practical guide to misleading scales, aggregates, and visual encodings in applied ML workflows.

misleading chartsapplied
Open guide
4

Respect the Data-Ink Ratio

Remove ink that does not explain data, structure, or context. Minimalism is useful when it protects attention, not when it removes necessary evidence.

Use in class

  • Remove 3D effects, heavy borders, redundant legends, and decorative backgrounds.
  • Keep scaffolding that improves reading: subtle gridlines, units, and reference values.
  • Let the data be visually louder than the furniture.
Ask: what can be removed without changing what a viewer can learn?

Examples

Tufte sparkline: tiny, dense, word-sized evidence

Tufte Sparklines

Sparklines compress trends into text without axes, boxes, or chart ceremony.

Tuftesparklines
Open Tufte note
Mike Bostock: Protovis sparklines

Protovis Sparklines

Working code examples show how much can be communicated with a few pixels and careful marks.

minimalinline
Open example
Stephen Few dashboard examples

Perceptual Edge Dashboard Design

Few's dashboard work emphasizes low-clutter monitoring, compact displays, and useful comparison.

dashboardclutter
Open PDF
5

Choose Strong Visual Encodings

Position on a common scale is usually easier to compare than angle, area, hue, or volume. Use weaker encodings only when they fit the question.

Use in class

  • Use scatterplots for relationships, dot plots for rank, bars for magnitude, and lines for time.
  • Avoid pie charts when many slices or close comparisons are needed.
  • Do not encode one variable in three redundant ways unless it improves accessibility.
Ask: what visual channel is carrying the most important number?

Examples

Financial Times chart-type vocabulary

FT Visual Vocabulary

Maps analytic intent to chart form, making encoding choices explicit rather than habitual.

chart choiceFT
Open visual vocabulary
Data Viz Catalogue: chart methods

Data Viz Catalogue

Chart-type references are useful when students need to match data structure to graphical form.

methodschart types
Open catalogue
Chartio guide to choosing charts

Chart Selection Guides

Useful as a classroom counterpoint: rules of thumb help, but the question still comes first.

selectionworkflow
Open guide
6

Show Variation, Not Only Averages

Averages are summaries, not evidence. Show distributions, outliers, clusters, sample size, and shape when they matter.

Use in class

  • Use dot plots, histograms, box plots, violins, and density plots to show spread.
  • Show raw points when the dataset is small enough.
  • Keep summary marks, but make them secondary to the pattern.
Ask: what important fact disappears if we replace this chart with a mean?

Examples

Anscombe quartet scatterplots

Anscombe's Quartet

The canonical lesson for why visualization belongs beside descriptive statistics.

variationdiagnostic
Open source
Datasaurus Dozen

Same Stats, Different Graphs

Matejka and Fitzmaurice extend Anscombe's lesson with many datasets sharing nearly identical summary statistics.

summary statsscatterplot
Open Autodesk paper page
Raincloud plots and distribution graphics

Raincloud Plots

Combines raw data, density, and summary statistics to avoid hiding distributional shape.

distributionraw data
Open article
7

Expose Mechanism and Causality

The strongest visualizations help viewers reason about why something happened, not only that it happened.

Use in class

  • Place the suspected cause and effect in the same visual field.
  • Use time, geography, annotation, and comparison to support causal reasoning.
  • Distinguish causal evidence from correlation and narrative convenience.
Ask: what mechanism does this chart make plausible, and what alternative remains?

Examples

John Snow cholera map

John Snow's Cholera Map

Deaths and pumps share a street map, turning a spatial cluster into an epidemiological argument.

mapmechanism
Open source
Minard Napoleon campaign map

Minard's Multivariate Explanation

Movement, loss, retreat, and temperature are arranged as one explanation rather than separate charts.

layeredexplanation
Open category
NASA Challenger O-ring case

Temperature as Mechanism

The key teaching move is to plot O-ring damage against temperature, the causal variable engineers worried about.

engineeringrisk
Open Stanford article
8

Use Small Multiples for Repeated Comparison

Repeated charts with the same scale let viewers compare many slices without decoding a tangled legend or overplotted mess.

Use in class

  • Keep the axes and design constant across panels.
  • Order panels by a meaningful variable.
  • Prefer small multiples over animation when side-by-side comparison matters.
Ask: would the viewer learn more if all panels were forced into one chart?

Examples

1870 Statistical Atlas of the United States

Statistical Atlas Small Multiples

Early census graphics use repeated structure to compare occupations, geography, and social measures.

historycensus
Open Commons category
Small multiples as frames of a movie

Tufte's Small Multiple Idea

A concise explanation of why repeated, comparable panels are powerful for trend and category comparison.

Tuftefacets
Open guide
Observable examples: small multiples in D3

Observable Plot Facets

Modern implementation examples show how faceting reduces overplotting while preserving comparison.

D3faceting
Open notebook
9

Layer and Separate Information

Use visual hierarchy to separate data, labels, gridlines, context, uncertainty, and annotations so the viewer sees the data first.

Use in class

  • Make data marks high contrast and supporting scaffolding low contrast.
  • Use whitespace and alignment instead of boxes when possible.
  • Highlight the few marks under discussion, not every mark.
Ask: what does the eye notice first, and is that the thing the chart is about?

Examples

John Snow cholera map

Snow's Map Layers

Street network, death marks, and pump locations each do different work without needing a separate chart.

layeringmap
Open original scan
Nightingale mortality diagram

Nightingale's Layered Causes

Color and area separate three causes of death while month and year remain visible.

layeringpublic health
Open source
Quartz Atlas chart style

Chart Style Guides

Newsroom style guides show how restrained grids, direct labels, and hierarchy create scan-friendly charts.

stylehierarchy
Open Chartbuilder
10

Integrate Words, Numbers, and Images

Clear annotation is not a crutch. Good labels, units, notes, and callouts help the viewer read evidence accurately.

Use in class

  • Use direct labels when a legend would force eye travel.
  • State units, denominators, and time windows close to the data.
  • Annotate surprising events, discontinuities, and data revisions.
Ask: could a viewer interpret this without reading a separate paragraph?

Examples

Minard Napoleon campaign map

Minard's Integrated Labels

Place names, temperatures, dates, and explanatory notes are embedded where they are needed.

annotationlabels
Open source
The Economist chart style guide

The Economist Style Guide

Direct labeling, concise titles, source notes, and restrained color are treated as part of the chart.

style guidelabels
Open PDF
Datawrapper annotation examples

Datawrapper Academy

Practical newsroom examples for title, subtitle, notes, color keys, and direct annotation.

annotationnewsroom
Open academy
11

Use Color Deliberately and Accessibly

Color should encode meaning, direct attention, and remain readable for people with color-vision deficiencies and in poor display conditions.

Use in class

  • Use sequential palettes for ordered data, diverging palettes around a meaningful midpoint, and qualitative palettes for categories.
  • Do not use rainbow scales for ordered data unless there is a strong domain reason.
  • Pair color with labels, position, shape, line style, or pattern when the distinction is important.
Ask: is the chart still readable in grayscale or by a colorblind student?

Examples

ColorBrewer map palettes

ColorBrewer

Cynthia Brewer's palette tool ties color choices to data type, print conditions, and colorblind safety.

palettemaps
Open ColorBrewer
Cividis and perceptual color maps

Cividis Colormap

A scientific color scale designed for perceptual uniformity and improved accessibility.

accessibilityscience
Open paper
Colorblind-friendly visualization guide

Multi-Channel Encoding

Examples and techniques for line styles, shapes, patterns, contrast, and redundant labeling.

accessibilitypractice
Open guide
12

Treat Maps as Arguments, Not Wallpaper

Maps are powerful because geography feels concrete. Use them when location matters, and choose projection, scale, and aggregation carefully.

Use in class

  • Use maps for spatial questions; use charts when ranking or exact comparison is the task.
  • Normalize counts when area or population size would mislead.
  • Consider cartograms, small multiples, or linked charts when geography hides the pattern.
Ask: what would this look like as a sorted bar chart, and which view is more honest?

Examples

John Snow cholera map

John Snow's Cholera Map

A map is justified because the theory is spatial: deaths cluster around a water pump.

spatialpublic health
Open source
Harry Beck's London Underground map

Tube Map Abstraction

Geographic accuracy is intentionally sacrificed to answer the rider's real question: how do I transfer?

abstractionnetwork
Open background
Worldmapper cartograms

Cartograms

Cartograms make the unit of comparison visible when land area would dominate attention.

cartogramnormalization
Open Worldmapper
13

Make Uncertainty Visible

A chart that hides uncertainty often looks more precise than the evidence allows. Show intervals, scenarios, sample sizes, revisions, and missingness when they affect interpretation.

Use in class

  • Use intervals, bands, distributions, ensembles, or simulation draws depending on the uncertainty type.
  • Use plain language to distinguish probability from vote share, risk from outcome, and model from measurement.
  • Show what is known, unknown, and unknowable.
Ask: what false certainty would a viewer take away if the uncertainty marks were removed?

Examples

FiveThirtyEight election forecast visualizations

Forecast Simulations

Dot swarms and scenario graphics help viewers understand many possible outcomes, not a single prediction.

probabilityforecast
Open showcase
NYT COVID uncertainty discussion

COVID Uncertainty

News graphics during the pandemic had to communicate ambiguity, missing data, and changing evidence.

uncertaintypublic data
Open Storybench article
Visualizing uncertainty research

Uncertainty Visualization

Research examples cover interval plots, gradient plots, quantile dotplots, and animated hypothetical outcomes.

researchintervals
Open overview
14

Use Scales That Match the Phenomenon

Linear, logarithmic, indexed, per-capita, and percent-change scales answer different questions. The scale must be explained and defensible.

Use in class

  • Use zero baselines for bar lengths unless there is a clear reason not to.
  • Use log scales for multiplicative growth, ratios, or wide ranges.
  • Normalize by population, exposure, or opportunity when raw counts are not comparable.
Ask: what question does this scale answer, and what question does it make harder?

Examples

Our World in Data: log scales and long-run trends

Log Scale for Growth

OWID often lets viewers switch scales, making the analytic consequences of scale visible.

log scaletime series
Open Grapher
COVID log-scale explainers

Exponential Growth

Log scales were central to explaining case growth rates, doubling times, and curve flattening.

exponentialpublic health
Open Datawrapper examples
Normalization in maps and rates

Rates vs Counts

Per-capita and rate charts prevent large-population places from dominating every comparison.

normalizationrates
Open choropleth guide
15

Make Interaction Earn Its Keep

Interaction should reveal detail, support exploration, or personalize the question. It should not hide the main argument behind unnecessary clicks.

Use in class

  • Give the static default view a clear takeaway.
  • Use hover, filters, and animation for secondary questions.
  • Use interaction to test a viewer's expectation when that expectation matters.
Ask: what does interaction let us learn that a static chart could not?

Examples

NYT You Draw It

Expectation First

Readers draw their guess before seeing the data, making learning active rather than passive.

interactionprediction
Open NYT example
Gapminder Tools

Animated Exploration

Time animation and country selection invite students to inspect both global structure and specific cases.

animationexploration
Open tools
Observable notebooks

Explorable Explanation

Observable examples make code, data, and chart behavior visible, which is useful for AI-assisted workflows.

notebookexplorable
Open D3 gallery
16

Show Data Provenance and Method

A serious chart tells viewers where the data came from, what was transformed, what was excluded, and how to reproduce or challenge the result.

Use in class

  • Include source, date accessed, definitions, and units.
  • Call out missing data, revisions, imputation, and filters.
  • Keep code or data links nearby when students should audit the work.
Ask: could another student recreate the chart from the information provided?

Examples

Our World in Data sources and downloads

Transparent Data Downloads

OWID charts expose source metadata, downloads, and chart configuration for reuse and audit.

provenancereuse
Open Grapher
FiveThirtyEight data repository

Public Data Behind Stories

Story data repositories make it possible to inspect assumptions and reproduce many published graphics.

datajournalism
Open data repository
NYT COVID-19 public data

Data Behind Public Tracking

A data repository can become part of the public evidence infrastructure, not merely a chart footnote.

public datamethod
Open repository

Source Notes

These links are included for classroom attribution and follow-up reading. Some examples are public-domain or openly licensed; others are linked as external teaching references.