data-visualization

Data Visualization

Intro

A good chart matches the chart type to the data relationship, annotates the insight rather than the raw point, and stays readable for colorblind viewers. The default fallbacks are bar (categorical) and scatter (continuous).

Overview

Chart type selection

Match chart to relationship, not to what looks impressive. Quick guide: comparison = bar, trend = line, distribution = histogram, relationship = scatter. When unsure, default to bar (categorical) or scatter (continuous). Never use a pie chart with more than four slices — use a horizontal bar instead. Avoid 3D charts entirely unless you are visualizing actual three-dimensional data. Full decision tree is in Level 3.

Color and accessibility

Use colorblind-friendly palettes: seaborn "colorblind", viridis, cividis. Never rely on color alone — add patterns, markers, or labels. Sequential data (low-to-high) needs a single-hue gradient or viridis. Diverging data (deviation from center) needs coolwarm or RdBu centered at zero. Categorical data needs distinct hues with sufficient contrast. Limit to seven or fewer colors; group or facet beyond that. Test with a colorblindness simulator before sharing.

Annotation

Annotate the insight, not just the data point. Always annotate max/min values that matter, threshold or target lines, and events that explain trend changes. Use direct labels instead of legends with 2–3 series. Arrows should point from label to data, not the reverse. Keep annotation text concise: "Peak: 1,247 (Q3 launch)" not "The maximum value occurred in Q3."

Dashboard layout

Place the most important metric in the top-left (reading order). Use a consistent grid: 2–3 columns, cards of equal height per row. Group related metrics visually. Filters live at the top, not buried in individual charts. Every dashboard needs a title, date range, and data freshness indicator. Limit to 6–8 visualizations per dashboard — more causes cognitive overload.

Storytelling

Lead with the conclusion: "Revenue grew 40% in Q3," not "Here is a chart of revenue." Structure as context → finding → implication. Use progressive disclosure: summary first, then supporting detail. Highlight the relevant; dim or remove the irrelevant. Annotate to guide the viewer's eye to the key insight.

Static vs interactive

Static charts (matplotlib, seaborn) for reports, papers, presentations. Interactive (plotly, Altair, D3) for dashboards and exploratory tools. Interactive needs tooltips, zoom for dense data, and filter controls. Static needs readability at print size (12pt+ fonts, 300 dpi). Always provide a static fallback for interactive charts.

Gotchas

Agent-specific failure modes — provider-neutral pause-and-self-check items:

• Truncating the bar chart y-axis to make differences look larger. A bar chart with a y-axis starting at 85% makes a 3-percentage-point difference look enormous. Bar charts must start at zero unless the baseline is explicitly labeled and well understood by the audience. Use a dot plot or connected line for before/after comparisons where zero is not meaningful.
• Dual y-axes implying a relationship that doesn't exist. Two series on dual axes can be made to look correlated or anti-correlated purely by scaling. The chart becomes a tool for misleading rather than informing. Use small multiples or index both series to a common baseline instead.
• Overplotting on scatter plots hiding the actual distribution. With more than a few hundred points, overlapping markers make density invisible — a dense cluster and a sparse one look identical. Use alpha transparency, 2D density (hexbin, KDE), or aggregate before plotting when the point count is large.
• Pie charts with more than four slices. Humans cannot accurately compare arc lengths or areas beyond four segments. Use a horizontal bar chart ordered by value; it is easier to read, easier to label, and easier to extend.
• Relying on color alone to encode information. Color-blind viewers (roughly 8% of men) cannot distinguish red from green or blue from yellow. Always pair color with a second encoding — position, shape, pattern, or direct label — so the chart is readable in grayscale.
• Default titles that describe the chart, not the finding. A title like "Monthly Revenue by Region" tells the viewer what data is shown but not what to conclude. A title like "Northeast revenue grew 34% while all other regions declined" tells them what to look for. Lead with the insight.
• Missing units on axes. "Revenue" on the y-axis without a currency and scale (thousands? millions?) forces the viewer to guess. Every axis must have a label that includes units. Every chart must specify the time range if time is involved.

Full reference

Chart selection decision tree

Comparison across categories:

• Vertical bar — magnitudes across < 15 categories. Avoid for time series (use line) or > 15 categories (use horizontal).
• Horizontal bar — long labels or > 10 categories. Avoid for time-based or continuous x-axis.
• Grouped bar — comparing 2–3 sub-groups within a category. Avoid > 3 sub-groups (facet instead) or when showing totals (stacked).

Trend over time:

• Line — continuous time axis, 1–5 series. Avoid with < 5 time points or > 7 lines.
• Area — emphasizing total volume or magnitude. Avoid for similar-magnitude series (lines occlude) or precise reading.

Distribution:

• Histogram — shape of one continuous variable.
• Box plot — comparing distributions across groups.
• Violin — like box plot but shows full shape; good for bimodal data.

Relationship:

• Scatter — two continuous variables. Use hexbin or 2D density for > 1000 points.
• Bubble — scatter with size as a third variable. Avoid for precise comparison (humans are bad at areas).
• Heatmap — magnitude across two categorical or ordinal dimensions. Avoid for sparse data or > 20 items per axis.

Composition (parts of a whole):

• Stacked bar — parts contributing to a total over time or category. Avoid > 5–6 segments.
• Pie — only with 2–4 slices and approximate proportions. Almost always replaceable with a horizontal bar.
• Treemap — hierarchical part-of-whole with many categories. Avoid when precise comparison matters.

Geographic:

• Choropleth — values for defined regions.
• Point/dot map — individual locations or events.

Relationship	First choice	Alternative	Avoid
Compare categories	Vertical bar	Horizontal bar, lollipop	Pie
Trend over time	Line	Area	Bar (unless few points)
Distribution	Histogram	Box, violin	Pie
Two variables	Scatter	Hexbin, 2D density	Line
Parts of whole	Stacked bar	Treemap	Pie (> 4 slices)
Correlation matrix	Heatmap	Clustered pairs	Table of numbers
Geographic	Choropleth	Point map	Non-geographic

Universal rules

1. Start with the simplest chart that answers the question.
2. If the legend has > 7 entries, redesign.
3. If a colleague cannot understand it in 10 seconds, simplify.
4. When comparing, use the same axis scale across panels.
5. Titles state the insight: "Sales doubled in Q3," not "Sales by Quarter."

Common mistakes

• Truncated y-axis on bar charts — bars must start at zero
• Dual y-axes — misleading; use two panels
• Overplotting — use alpha, jitter, hexbin, or density
• Missing units on axes
• Default titles ("Figure 1") instead of the finding
• Legend far from the data it describes
• Too many decimal places — round to meaningful precision

Export checklist

Before sharing, verify: title states the finding; both axes labeled with units; legend present or direct labels used; colorblind-friendly palette; 300 dpi for print, SVG/PDF for scalable; fonts readable at final display size.

Worked examples

• 5 product lines × 24 months: Multi-line chart, one line per product, distinct colorblind-safe colors with direct end-of-line labels instead of a legend. Title states the trend ("Product A overtook B in mid-2025"), annotate the crossover point, light gridlines on white background.
• 15-chart dashboard is overwhelming: Audit for redundancy, group remaining charts by domain, move detail charts to a drill-down page, keep 6 key metrics on the main view, add a summary KPI row with sparklines at the top.
• Overplotted scatter: Reduce marker size, set alpha=0.3, switch dense regions to hexbin or 2D density, add marginal histograms, label outlier clusters.