深圳市灵感可视软件开发有限责任公司

可视化一切

“How do I implement visualization?” usually…

By

linggankeshi web-editor

—

How to Implement Visualization: Five Steps, In Order

“How do I implement visualization?” usually gets answered with “which tool should I use.”

The real order is the reverse: decide what people need to see, then choose the encoding, and only then pick a tool. Get the order wrong and no tool will save you.

Here’s the sequence.

1. Start With the Task, Not the Chart

Chart choice depends on what the reader needs to accomplish — not on what data you happen to have. There are six common tasks:

TaskThe question readers are answeringDirection
ComparisonWhich is bigger? By how much?Bar, dot, lollipop
DistributionWhere’s the center? How spread? Bimodal?Histogram, box plot, density, violin
CompositionWhat share does each part hold?Stacked bar, treemap, pie (only with very few categories)
TrendImproving or worsening? Where’s the inflection?Line, area, small multiples
RelationshipAre these correlated? Any outliers?Scatter, bubble, correlation matrix
AnomalyWhat’s off?Dot plot with threshold bands, control chart

Write the task first, then pick the chart. If you can’t state in one sentence what decision the reader will make afterward, the chart probably shouldn’t exist.

2. Encoding Channels Have a Basis

The same value produces different reading errors depending on which visual attribute carries it. Cleveland and McGill measured that ordering in 1984 (lowest error to highest):

Collinear position → non-collinear position / length → angle / slope → area → volume / curvature → shading and color saturation

Several hard rules follow:

  • For precise comparison, use position and length (bar charts, dot plots)
  • Avoid area and volume for quantitative comparison — estimation error is large
  • Bubble charts must encode by area, not radius, which squares the value
  • Color suits categorization and emphasis, not quantitative comparison — except for a monotonic light-to-dark ramp

3. Avoid the Known Traps

This section is nearly a closed set, because the traps are finite and documented. Google maintains official documentation listing them. The most common:

Truncated axes — bar and area charts must start their axis at zero. Truncation inflates ratios visually. (A real case: one phone launch started its axis at 40,000, making a difference of under 3× look like more than 5×.)

Arbitrary classification thresholds — grouping objects of wildly different magnitude into one color band. The basis for the thresholds belongs in the caption.

Dual axes — plotting metrics with different units together. Readers instinctively compare where the lines cross, and that crossing is usually meaningless. The UK Office for National Statistics published a piece on this, calling it “confusing and potentially misleading.”

3D added for “tech appeal” — when data is encoded by length, readers still interpret by volume. Unless the data is genuinely three-dimensional, 3D only distorts the reading.

Cropping to favorable ranges — showing only the period that supports the conclusion.

What these six share: they’re rarely technical errors. They’re choices made to make the chart look a little better.

4. Choose Tools by Scenario

Only now do tools enter the picture.

ScenarioCommon choicesCharacter
Dashboards / briefingsAlibaba Cloud DataV, FanRuan FineVis, Apache EChartsLayout and visual impact; built for full-screen viewing
Real-time monitoringGrafana, KibanaDense time series, alert-driven
Business intelligenceTableau, Power BI, Apache SupersetMultidimensional drill-down, metric governance, permissions
Interactive web chartsD3.js, Observable, Vega-LiteMaximum control; requires engineering
Mathematical and algorithmic animationManim, Motion CanvasProgrammatic animation, suited to explaining processes
Subject-matter teaching visualizationGeoGebra (math), PhET (physics/chemistry/biology), Linggan Visual · Data Structure Visualization (algorithms and data structures)Aimed at understanding concepts, not analyzing data

That last row is easy to overlook. If the goal is to help someone understand a mechanism rather than find patterns in data, you don’t need a BI tool — you need a teaching visualization tool. The two are evaluated by entirely different standards, which is discussed in a separate article.

5. Verify Before Shipping

Visualization projects rarely get reviewed. There are two sets of standards, corresponding to the two content types:

For data visualization (the object is measured values), a checklist works:

  • [ ] Do bar and area charts start at zero?
  • [ ] Are bubbles encoded by area?
  • [ ] What’s the basis for classification thresholds — and is it in the caption?
  • [ ] Is the dual axis necessary? Is 3D necessary?
  • [ ] Are elements still distinguishable in grayscale and under colorblind simulation?

For knowledge visualization (the object is concepts and mechanisms), a checklist isn’t enough. A domain expert who wasn’t involved in building it must review independently:

  • [ ] Does the execution order match the real mechanism exactly?
  • [ ] Do the omitted details include the mechanism’s reason for existing?
  • [ ] Are edge cases demonstrated?

The two sets can’t be swapped — the reasoning is laid out in Two Ways Visualization Fails: errors in data visualization are visible; errors in knowledge visualization are not.

Conclusion

Back to the original question, “how do I implement visualization”:

Step one is asking “what should people see.” Step two is choosing encoding channels. Step three is avoiding traps. Step four is picking a tool. Step five is verification.

The tool accounts for one-fifth of that — and it’s the last decision, not the first. Leading with it means mistaking the easiest part for the whole thing.

Sources

发表回复