Why we draw numbers
A table of figures is precise, but it is hard to read quickly. Our eyes are good at spotting differences in length, position and slope, and much slower at comparing rows of digits. A chart turns numbers into shapes that the brain can compare almost instantly. That is the basic promise of data visualization: make patterns, gaps and outliers visible.
Take a hypothetical list of monthly website visits: 12, 14, 13, 15, 18, 30, 17, 16 (in thousands). You can find the 30 by scanning, but on a line chart it stands out immediately as a spike. Whether it is a genuine surge or an error is a separate question, and this is one place where data cleaning matters before drawing anything.
Matching the chart to the question
Different charts answer different questions. A common rule is to begin with what you want the reader to see.
- Comparing categories: a bar chart. Bars start at a common baseline, which makes lengths easy to compare fairly.
- Change over time: a line chart. The slope shows rising or falling, and gaps between points show speed.
- Relationship between two measurements: a scatter plot. Each dot is one item, and the cloud’s shape suggests whether they move together. Remember that a visible pattern is not proof of cause, a theme in the discussion of correlation versus causation.
- Distribution of values: a histogram, which groups values into ranges and shows how many fall in each.
- Parts of a whole: a stacked bar or, with only a few slices, a pie chart. Pies become hard to read with many slices or similar sizes.
- Two dimensions at once: a heatmap, using color intensity to show size across a grid, such as hours of the day against days of the week.
Maps are best when location truly matters. Otherwise, a simple bar chart sorted by value is often clearer than a map.
Design choices that help
Good charts tend to share a few habits.
- A clear title that states the point or the subject, plus labeled axes with units.
- Sorted categories, unless there is a natural order such as months.
- Limited color, used to highlight what matters instead of decorating everything. Color-blind-friendly palettes help people who cannot tell certain colors apart, and labeling directly can avoid relying on color at all.
- Little clutter. Heavy gridlines, 3D effects and decorative backgrounds make comparison harder.
- Direct labels or legends placed near the data, so readers do not have to look back and forth.
- A source and date, so readers can judge how current the figures are.
How charts mislead
Charts can mislead without anyone intending it. Learning the common patterns makes you a better reader.
Truncated axes. If a bar chart’s vertical axis starts at 95 instead of 0, a change from 96 to 99 looks enormous. Bars encode value by length, so cutting the baseline distorts the comparison. Line charts have more freedom to zoom in, but the axis range should still be visible and honest about it.
Stretched or squashed proportions. The same line can look flat or dramatic depending on how tall the chart is relative to its width.
Cherry-picked time windows. Starting a line chart at a low point makes growth look steady; starting at a high point makes the same series look like a decline. Ask what happened before and after the period shown.
Double axes. Two lines with separate vertical scales can be made to look as though they track each other closely or diverge sharply, just by changing the scales.
Misleading areas and icons. If a circle or picture is made twice as tall, its area becomes four times as large, which exaggerates the difference. Sizes should reflect the values in a way that readers can judge.
Omitting context. A figure that doubled sounds dramatic, but from what base? Going from 1 to 2 is a doubling too. Showing absolute numbers beside percentages helps. The same goes for averages, which can hide very different groups.
Too many categories or lines. When ten lines overlap, nothing is clear. Highlighting one or two and graying out the rest usually works better.
Visualization in analysis work
Charts are not only for final reports. Analysts often plot data early in a project to spot problems: values that are impossible, gaps from a broken sensor, or a distribution that is not what was assumed. Dashboards used for data analytics bring several charts together so teams can follow key measures at a glance, and they depend on the underlying data being timely and correct. Many dashboards pull from a data warehouse built to bring numbers from several systems together for reporting.
In machine learning, plots of how well a model performs over time or against its training data help people notice problems such as model overfitting, where a model fits its examples too closely and does poorly on new cases.
A short checklist for reading any chart
- What exactly is being measured, and in what units?
- Where does each axis start, and is the scale uniform?
- What period or group is shown, and what is left out?
- Are these counts, percentages or averages, and of what?
- Who made the chart, and what might they want you to conclude?
- Does the visual impression match the actual numbers?
Telling the story honestly
The best charts do not shout. They show the data plainly enough that a reader can reach a sound conclusion, including a conclusion the author did not expect. A chart’s job is to make the numbers easier to understand, not to make an argument look stronger than the evidence supports.