Data Visualization

Overview

Match the chart to the data relationship, encode quantities in perceptually accurate channels, avoid distortion, and make it accessible. This skill is the design layer above the plotting library.

Chart selection by data relationship

Perceptual accuracy (why bars beat pies)

Cleveland–McGill ranking of how accurately humans decode a quantity:

position on common scale > position on non-aligned scale > length > angle/slope > area > volume > color hue/saturation.

• Encode the most important quantity in position/length (bar, dot, line), not area or color.
• Pie/donut = angle+area (weak); bubble = area (people underestimate large circles — area scales as r², so double the value ≠ double the radius). Reserve area/color for secondary dimensions.
• Sort categorical bars by value (not alphabetically) unless order is semantic — sorting is the insight.

Avoiding distortion (the "lie factor")

• Truncated y-axis exaggerates change. Bar charts must start at 0 (bar length encodes the value). Line/time-series may zoom the axis to show variation — but label it clearly; don't imply a 2% change is a cliff.
• Dual y-axes manufacture spurious correlation and let you slide two series' scales arbitrarily. Replace with indexed series (all = 100 at t0), two small multiples, or a ratio.
• Aim for lie factor ≈ 1 (graphic effect size / data effect size). Don't use 3D, shadows, or area to represent 1D quantities.
• Chartjunk: maximize data-ink ratio (Tufte) — drop gridline clutter, heavy borders, redundant legends, background gradients. Direct-label lines instead of a legend when few series.

Color and accessibility

• Sequential: viridis/cividis — perceptually uniform and colorblind-safe. Avoid jet/rainbow (non-uniform; invents false boundaries; not CVD-safe).
• Diverging (signed around a midpoint): RdBu, coolwarm — set the neutral point at the meaningful zero.
• Qualitative (categories, ≤~8): Okabe–Ito or ColorBrewer Set2. Beyond ~8 colors, hue stops being distinguishable — use small multiples, top-N+Other, or direct labels instead.
• Never rely on hue alone (~8% of men have CVD): add redundant encoding — shape/linestyle, direct labels, or patterns. Red/green is the worst offender.
• WCAG AA: 4.5:1 contrast for text, 3:1 for meaningful graphics. Provide alt text and a data-table fallback for interactive charts; support keyboard nav.

Handling data volume

• Overplotting: for dense scatter use hexbin / 2D density / alpha-blending / sampling — not 100k opaque points. For many time series use small multiples or highlight-one-fade-rest, not a spaghetti chart.
• High-cardinality categoricals: top-N by value + an "Other" bucket; horizontal sorted bars; never a 30-slice pie or 30-color legend.
• Time-series downsampling: don't push 1M points into 800px. Aggregate to the display resolution (LTTB — largest-triangle-three-buckets — preserves visual shape) or roll up to buckets. Show p50/p95/p99, not just the mean, for latency; show gaps for missing data rather than interpolating across them.
• Log scale for data spanning orders of magnitude or multiplicative/growth relationships — label it explicitly (readers assume linear). ⚠ Log can't show zero/negative values.

Library selection

Rule of thumb: standard statistical chart → Vega-Lite/Plotly/Seaborn; bespoke/interactive-novel → D3/Observable Plot; ops time-series → Grafana. Reach for D3 only when a grammar-of-graphics tool genuinely can't express the design.

Dashboard design

• Most-important-top-left (F/Z reading pattern); lead with 4–6 KPIs (value + trend/delta + context/target), details below, progressive disclosure via drill-down.
• Consistent axes, color meaning, and units across panels — the same color must mean the same thing everywhere.
• Provide context on every number: baseline, target, or prior period. A lone "1,240" is noise.
• Analytics: date-range + filters; export (CSV/PNG); cache expensive queries.
• Monitoring: time-series lines; threshold bands; current status prominent; percentiles not averages; auto-refresh 30–60s.
• ML: train+val curves on one axis (gap = overfit); normalized confusion matrix; horizontal feature-importance bars; per-experiment comparison.

Reference examples

Matplotlib multi-panel with target band and value labels:

import matplotlib.pyplot as plt
import pandas as pd

df = pd.DataFrame({
    "date": pd.date_range("2024-01-01", periods=12, freq="ME"),
    "revenue": [100,120,115,140,155,170,165,180,195,210,225,250],
    "target":  [110,115,120,130,145,160,175,185,200,215,230,245],
})
fig, ax = plt.subplots(figsize=(8, 5))
ax.plot(df.date, df.revenue, marker="o", lw=2, label="Actual")
ax.plot(df.date, df.target, ls="--", lw=2, label="Target")
ax.fill_between(df.date, df.revenue, df.target,
                where=df.revenue >= df.target, alpha=0.25, color="#029E73")  # ahead
ax.fill_between(df.date, df.revenue, df.target,
                where=df.revenue <  df.target, alpha=0.25, color="#D55E00")  # behind
ax.set(title="Monthly Revenue vs Target", xlabel="Month", ylabel="Revenue ($K)")
ax.legend(); ax.margins(x=0.01)
plt.tight_layout()

Interactive Plotly time series with range selector + moving average:

import plotly.graph_objects as go
fig = go.Figure()
fig.add_scatter(x=df.date, y=df.value, mode="lines", name="Daily",
                hovertemplate="%{x|%b %d}<br>%{y:.1f}<extra></extra>")
fig.add_scatter(x=df.date, y=df.value.rolling(7).mean(), mode="lines",
                name="7d MA", line=dict(dash="dash"))
fig.update_layout(template="plotly_white", hovermode="x unified",
                  xaxis=dict(rangeslider=dict(visible=True)))

Colorblind-safe qualitative palette (Okabe–Ito) with redundant value labels:

OKABE_ITO = ["#E69F00", "#56B4E9", "#009E73", "#F0E442", "#0072B2", "#D55E00", "#CC79A7"]
bars = ax.bar(labels, values, color=OKABE_ITO[:len(values)])
ax.bar_label(bars)  # direct labels: readable without color perception

Gotchas

• matplotlib freq="M" is deprecated → use "ME" (month-end) in pandas ≥2.2.
• Seaborn style names are versioned: plt.style.use("seaborn-v0_8-whitegrid") (the bare seaborn-* names were removed in Matplotlib 3.6+).
• Pie charts of percentages that don't sum to 100, or with negatives, are meaningless — validate the data is a true part-to-whole first.
• Rainbow/jet colormaps create perceptual banding that fabricates structure — banned for quantitative encoding.
• Averaging percentiles or averaging rates across groups (Simpson's paradox) misleads — aggregate raw counts, then compute the ratio.

Checklist — before shipping a chart/dashboard

• Chart type matches the data relationship (not just what's easy)
• Key quantity encoded in position/length, not area/color alone
• Bars start at 0; no deceptive dual axes; lie factor ≈ 1
• Colorblind-safe palette; meaning not conveyed by hue alone; 4.5:1 text contrast
• Axes labeled with units; log scale (if used) labeled; legend or direct labels present
• Dense data downsampled/aggregated; high-cardinality categories bucketed (top-N + Other)
• Context shown (baseline/target/prior); percentiles not just means for latency
• Alt text / data-table fallback for accessibility; consistent color meaning across panels
• Validated against real production-scale data; dashboard loads < 3s