Seaborn is built on top of Matplotlib and specializes in statistical visualization — it's the library you'll use most during EDA, because it makes distribution, correlation and category-comparison plots a single line of code.
The EDA Workhorses
import seaborn as sns
import matplotlib.pyplot as plt
# 1. Correlation heatmap — spot multicollinearity and predictive features fast
corr = df.corr(numeric_only=True)
sns.heatmap(corr, annot=True, cmap="coolwarm", fmt=".2f")
plt.show()
# 2. Boxplot — compare a numeric feature across categories, spot outliers
sns.boxplot(x="department", y="salary", data=df)
plt.show()
# 3. Pairplot — every numeric feature plotted against every other, at a glance
sns.pairplot(df[["age", "income", "credit_score", "churned"]], hue="churned")
plt.show()
# 4. Distribution plot with a smoothed curve
sns.histplot(df["income"], kde=True)
plt.show()Reading a Correlation Heatmap
Values range from -1 to +1: near +1 means two features move together strongly, near -1 means they move in opposite directions, and near 0 means little linear relationship. Two features that are highly correlated with each other (not with the target) is a signal for multicollinearity — worth addressing in Feature Selection. A feature highly correlated with the target is a promising predictor.
Why a Pairplot with hue Is So Useful
Coloring points by the target class (hue="churned") instantly shows whether a feature — or a combination of two features — visually separates the classes. If two classes overlap completely in every pairwise plot, a linear model will likely struggle, hinting you may need feature engineering or a non-linear model.
Practical Use Cases
- Spotting multicollinearity between features before modeling
- Comparing a numeric feature's distribution across categories or classes
- Confusion matrix visualization after evaluating a classifier
Common Mistakes
- Interpreting correlation as causation — a strong correlation in a heatmap never proves one feature causes another.
- Running
df.corr()only on numeric columns without first checking which columns are actually numeric vs. mistakenly-numeric categorical codes.
Interview Relevance
Q: "How would you check for multicollinearity before training a linear model?" Plot a correlation heatmap between features (not the target) — pairs with very high correlation (e.g. >0.9) are candidates to drop or combine, since they can destabilize a linear model's coefficients.
Practice Question
A correlation heatmap shows square_feet and num_rooms have a correlation of 0.93 with each other. What issue could this cause for a linear regression model, and what would you do about it?
