Heatmaps

A heatmap is a grid of colored cells, where each cell's color encodes a numeric value. It's the natural chart for any data that sits in a two-dimensional table — categories on rows and columns, numbers in the cells.

You will see heatmaps used for correlation matrices, schedule calendars, hourly traffic, missing-data maps, and confusion matrices in ML — all the same chart, applied to different data shapes.

When a heatmap is right

Use a heatmap when:

• You have a 2-D table of values: rows × columns of numbers.
• You want to spot patterns, clusters, or anomalies across both axes simultaneously.
• The number of rows and columns isn't huge — a 20×20 grid is comfortable, a 1000×1000 grid is a different kind of chart.

A simple heatmap

Plotly Express offers px.imshow (for image-like 2-D matrices) and px.density_heatmap (which bins two continuous variables on the fly).

Where the scatter plot showed every dot, this shows cell counts. The hot cells (yellow) are where most diners fall — small-to-medium bill, small tip.

A correlation matrix heatmap

A classic use is a correlation matrix — every numeric column correlated against every other.

Three Plotly Express tricks worth noticing:

• text_auto=True writes the numeric value in each cell.
• color_continuous_scale="RdBu_r" picks a diverging palette.
• color_continuous_midpoint=0 centers the white at 0 so positive correlations are blue and negative are red.

For diverging data (anything with a meaningful zero, like correlations or year-over-year change), always use a diverging palette centered at zero.

A calendar/schedule heatmap

Heatmaps also show counts across two categorical axes:

The hot cell (Sat dinner) leaps out immediately. A table of the same numbers would require the reader to scan; the heatmap is read in a single glance.

Choosing a color scale

This is the single most important decision when making a heatmap.

• Sequential (light → dark of one hue): "Viridis", "Plasma", "Blues", "YlOrRd". Use for values that go from low to high with no special midpoint.
• Diverging (color → white → other color): "RdBu", "RdBu_r" (reversed), "PiYG". Use for values with a meaningful zero (correlations, deviations, gain/loss).
• Avoid Jet / rainbow. It is perceptually non-uniform — bright cyan and yellow bands look like ridges in the data even where there aren't any. Viridis is its modern replacement.

Pitfalls of heatmaps

• Too many cells (1000×1000) make individual cells invisible. Either aggregate or switch to a density-based representation.
• Wrong palette (rainbow / Jet) introduces false structure.
• Forgetting the colorbar legend — without it, the chart's colors mean nothing to the reader. Always show the legend.
• Truncating the color range can mask outliers, similar to a truncated y-axis on a bar chart.

Check your understanding

QuestionSelect one

A heatmap is best suited for visualizing:

A single numeric variable's distribution.

A trend over time.

A 2-D table of numeric values — rows × columns — where you want to spot patterns or hotspots.

A part-of-a-whole composition.

QuestionSelect one

For a correlation matrix heatmap (values from -1 to +1), which color palette is most appropriate?

A sequential palette like Viridis.

A categorical palette like Plotly.

A diverging palette like RdBu or RdBu_r, centered at zero.

A grayscale ramp.

QuestionSelect one

Why should you avoid the Jet (rainbow) colormap for heatmaps?

It is copyrighted.

It is only available on Windows.

It is perceptually non-uniform — bright cyan and yellow bands appear as visual "ridges" even where data is flat, leading viewers to see structure that isn't there.

It uses too many colors.