EDA Visualizations Distributions And Relationships

Why this matters

As a Data Scientist, you constantly answer questions like: Is the target skewed? Which features correlate with each other? Are there segments behaving differently? Fast, clear visual explorations of distributions and relationships help you catch data issues early, find signals for modeling, and explain insights to teammates and stakeholders.

• Spot data quality issues: impossible values, heavy tails, multimodal shapes.
• Guide feature engineering: transformations (log), binning, interaction terms.
• De-risk models: find leakage, confounders, or dominant segments.
• Communicate findings: visuals are memorable and actionable.

Who this is for

• Current and aspiring Data Scientists who need reliable EDA habits.
• Analysts and ML engineers validating features and datasets.
• Anyone preparing data stories for product, ops, or leadership.

Prerequisites

• Comfort with basic statistics: mean, median, variance, percentiles.
• Basic plotting in Python (pandas, seaborn, matplotlib) or a similar tool.
• Ability to load and clean data (missing values, types).

Learning path

1. Master univariate distributions (histogram, KDE, ECDF, box/violin).
2. Explore relationships between two variables (scatter, hexbin, 2D KDE).
3. Compare groups (box/violin/ridge, small multiples, faceting).
4. Handle scale, skew, and overplotting (log scales, bin width, alpha).
5. Build a repeatable EDA workflow with notes and saved figures.

Concept explained simply

Distribution shows how values are spread for one variable. Relationship shows how two variables move together. We visualize distributions to see shape, center, spread, and outliers. We visualize relationships to detect trends, clusters, and interactions.

Core plot types you will use 90% of the time

Univariate (numeric)

• Histogram: count of values in bins. Control bin width for clarity.
• Density (KDE): smoothed curve of the distribution; good for comparisons.
• ECDF: cumulative proportion up to each value; great for tails and medians.
• Box/Violin: compact summary across categories; violin shows shape.

Categorical counts

• Bar chart: counts or proportions per category. Prefer proportions when group sizes differ widely.

Two numeric variables

• Scatter plot: the default; add transparency (alpha) and jitter if overlapping.
• Hexbin/2D density: for large datasets with heavy overlap.
• Line plot: if x is ordered (time or sequence).

Numeric vs categorical

• Box, violin, ridge plots: compare distributions across categories.
• Strip/Swarm: show individual points for small datasets.

Multivariate

• Colored/Faceted scatter: segment by hue or small multiples.
• Pair plot: quick overview of many numeric relationships at once.

Choosing the right plot (decision guide)

1. One variable? Use histogram/KDE/ECDF. Need compact summary across groups? Box/violin.
2. Two numeric variables? Start scatter; if overplotting, use alpha or hexbin.
3. One numeric, one categorical? Box/violin; add swarm for sample-size context.
4. Many variables? Pair plot for quick screening, then focus on key pairs.
5. Skewed or wide-range data? Consider log scales before interpreting.

Worked examples

Example 1: Customer spend is skewed

Goal: understand the distribution of monthly_spend to decide on transformation.

# Python (pandas + seaborn)
import pandas as pd, seaborn as sns, matplotlib.pyplot as plt
sns.set(style="whitegrid")
df = pd.read_csv("customers.csv")
fig, axes = plt.subplots(1, 3, figsize=(13,4))
# Histogram
sns.histplot(df["monthly_spend"], bins=30, ax=axes[0])
axes[0].set_title("Histogram: monthly_spend")
# ECDF
sns.ecdfplot(df["monthly_spend"], ax=axes[1])
axes[1].set_title("ECDF: monthly_spend")
# Log scale
sns.histplot(df["monthly_spend"], bins=30, ax=axes[2])
axes[2].set_xscale('log')
axes[2].set_title("Histogram (log x)")
plt.tight_layout()

• Interpretation: heavy right tail; top 10% contribute most revenue.
• Action: use log(monthly_spend + 1) for modeling to stabilize variance.

Example 2: Relationship with overplotting

Goal: visualize price vs. size for housing; handle 80k points.

fig, axes = plt.subplots(1, 2, figsize=(12,4))
# Scatter with alpha
sns.scatterplot(data=df, x="size_sqft", y="price", alpha=0.2, ax=axes[0])
axes[0].set_title("Scatter (alpha=0.2)")
# Hexbin (matplotlib)
axes[1].hexbin(df["size_sqft"], df["price"], gridsize=40, cmap="viridis")
axes[1].set_title("Hexbin")
plt.tight_layout()

• Interpretation: positive association; hexbin reveals dense band.
• Action: consider log scales; add region as hue or facet to check confounding.

Example 3: Numeric vs categorical (delivery time by region)

plt.figure(figsize=(9,4))
sns.violinplot(data=df, x="region", y="delivery_days", inner=None)
sns.boxplot(data=df, x="region", y="delivery_days", width=0.15, color="white")
sns.stripplot(data=df, x="region", y="delivery_days", color="black", alpha=0.3, size=2)
plt.title("Delivery days by region")

• Interpretation: Region South has wider spread and higher median.
• Action: Investigate carriers in South; consider segmenting models by region.

Practical EDA workflow you can reuse

1. Load and preview data: types, missingness, quick describe().
2. Plot univariate distributions for key numeric and categorical variables.
3. Check relationships: target vs features; feature vs feature for leakage.
4. Segment by important categories (hue or facet) to reveal hidden patterns.
5. Adjust scales and reduce clutter (log axes, alpha, bin widths); annotate observations.

Exercises (hands-on)

Do these now. They mirror the graded exercises below so you can compare your work.

Exercise 1: Retail orders EDA

Dataset columns: order_id, customer_id, order_value, items_count, is_expedited, region, order_date.

• Plot the distribution of order_value. Identify skew and suggest a transformation if needed.
• Visualize the relationship between items_count and order_value. Mitigate overplotting if necessary.
• Compare order_value across regions using an appropriate plot.

import seaborn as sns, matplotlib.pyplot as plt
# 1) Distribution
sns.histplot(df["order_value"], bins=30)
plt.figure(); sns.ecdfplot(df["order_value"])  # check tails
# If heavy right skew, use log
plt.figure(); sns.histplot(df["order_value"], bins=30); plt.xscale('log')
# 2) Relationship
plt.figure(figsize=(6,4))
sns.scatterplot(data=df, x="items_count", y="order_value", alpha=0.4)
# 3) Across regions
plt.figure(figsize=(8,4))
sns.violinplot(data=df, x="region", y="order_value", inner=None)
sns.boxplot(data=df, x="region", y="order_value", width=0.15, color="white")
  

Exercise 2: SaaS metrics EDA

Dataset columns: account_id, monthly_active_users, sessions_per_user, plan_tier, churned, revenue_mrr.

• Plot revenue_mrr distribution and compare by plan_tier.
• Check relationship: monthly_active_users vs revenue_mrr; handle outliers and scale issues.
• Segment the scatter by churned to see if patterns differ.

# Distribution by plan
sns.violinplot(data=df, x="plan_tier", y="revenue_mrr", inner=None)
sns.boxplot(data=df, x="plan_tier", y="revenue_mrr", width=0.15, color="white")
# Scatter with log scales
plt.figure(figsize=(6,4))
sns.scatterplot(data=df, x="monthly_active_users", y="revenue_mrr", alpha=0.3, hue="churned")
plt.xscale('log'); plt.yscale('log')
# Optional smoother
sns.lmplot(data=df, x="monthly_active_users", y="revenue_mrr", hue="churned", lowess=True)
  

• Checklist to self-grade:
  • You chose plots that match the variable types.
  • You addressed skew/scale and overplotting.
  • You wrote 1–2 takeaways for each plot.

Common mistakes and how to self-check

• Too-wide or too-narrow histogram bins. Fix: tune bins until shape is readable.
• Overplotting hides structure. Fix: alpha/jitter or switch to hexbin/2D density.
• Comparing raw counts across unequal groups. Fix: use proportions, box/violin, or ECDF.
• Ignoring scale/skew. Fix: try log axes; compare both raw and transformed.
• Trend line misuse. Fix: use LOWESS for nonlinear; don’t force linear fits.
• Bar charts not starting at zero. Fix: always start at zero for bar lengths.
• Cherry-picking views. Fix: show at least one univariate and one bivariate view per key variable.

Practical projects

• Product A/B results audit: visualize metric distributions pre/post, use ECDF for uplift tails, and segment by cohort.
• Credit risk profile: compare income and utilization distributions by default status; use hexbin for joint patterns.
• Sensor reliability: plot time-to-failure distributions by device model; investigate temperature vs failure with LOWESS.

Quick test

You can take the quick test below to check mastery. Everyone can take it for free. Progress is saved only for logged-in users.

Next steps

• Practice on a dataset you know well; summarize 3–5 insights with plots.
• Create a reusable notebook template for univariate, bivariate, and segmented EDA.
• Move on to multivariate and time-based visualizations; integrate results into dashboards.

Mini challenge

Pick one numeric target and three candidate features from any dataset. In one notebook page: (1) show the target distribution (raw and transformed if needed), (2) show each feature’s relationship to the target (choose the right plot), (3) write 5 bullet insights. Keep it under 15 minutes.