[reference] — from the roadmap.sh data-analyst/ai-data-scientist roadmaps. Two jobs: exploring data (charts for yourself, during EDA) and communicating findings (charts for others, the deliverable). Getting these right is a large part of a data scientist’s actual impact.

Choosing the right chart

The most common mistake is picking a chart type by habit rather than by what you’re trying to show. Match the chart to the question:

You want to show…Use
Comparison across categoriesbar chart (horizontal if many/long labels)
Trend over timeline chart
Distribution of one variablehistogram, box plot, violin plot
Relationship between two numericsscatter plot
Correlation across many variablesheatmap
Composition (parts of a whole)stacked bar — rarely a pie chart
Drop-off through stagesfunnel chart

Some hard-won specifics: pie charts are usually a bad choice (humans compare angles poorly — a bar chart almost always reads better); dual y-axes mislead (they let you imply a relationship by arbitrary scaling); and for distributions, a box or violin plot shows spread and outliers a bar-of-averages hides.

The libraries and tools

  • Matplotlib — the Python foundation; total control, verbose. Everything else sits on it.
  • Seaborn — statistical plotting on top of Matplotlib; beautiful defaults, one-liners for common EDA charts. The everyday EDA tool.
  • Plotly / Bokeh / Altair — interactive charts (hover, zoom), good for dashboards and notebooks.
  • ggplot2 (R) — the “grammar of graphics” reference implementation; many data scientists prefer R specifically for it.
  • BI tools — Tableau, Power BI, Looker: drag-and-drop dashboards for business stakeholders, no code. Even Excel (pivot tables, charts) is a real reporting tool for analysts. These are how findings reach non-technical decision-makers.
import seaborn as sns
sns.histplot(df["income"])                  # distribution
sns.scatterplot(data=df, x="age", y="spend") # relationship
sns.heatmap(df.corr(), annot=True)           # correlation matrix

The grammar of graphics

The idea (from Leland Wilkinson, realized in ggplot2 and Altair) that any chart decomposes into layers: data → mappings of variables to visual channels (x, y, color, size, shape) → geometry (points, bars, lines) → scales and facets. Thinking this way — “which variable maps to which channel?” — is more powerful than memorizing chart types, and it’s why faceting (small multiples: the same chart repeated per subgroup) is often clearer than cramming everything into one busy plot.

Communicating honestly

Visualization is uniquely easy to mislead with — sometimes accidentally, sometimes not. The integrity rules:

  • Don’t truncate the y-axis to exaggerate a difference (start bar charts at zero). A tiny change made to look huge is the most common deceptive chart.
  • Avoid chartjunk — 3D effects, heavy gridlines, decoration. Maximize the “data-ink ratio” (Tufte): every mark should carry information.
  • Label directly, minimize legends, and let the title state the takeaway (“Conversion rose 3% after launch”), not just the variables (“Conversion over time”).
  • Use color intentionally — for meaning, not decoration; keep palettes colorblind-safe (~8% of men are red-green colorblind), and use sequential/diverging palettes appropriately (never a rainbow for ordered data).
  • Show uncertainty — error bars / confidence bands, so a noisy estimate isn’t read as exact.

The goal is the same as the whole role: help someone reach a correct conclusion quickly. A chart that’s pretty but misleading is worse than no chart.

↑ Back to 01 — Data Scientist