Visualization for Class Imbalance — The Little Class That Could (and Often Can't)

"If your positive class is rarer than a unicorn sighting, you're not doing EDA — you're performing archeology."

You're arriving at the party having already seen how to visualize regression targets and pairwise relationships. Great — you've got context. You also just finished wrangling and feature-engineering your dataset, so features are clean, encoded, and not leaking like a sieve. Now the last obnoxious guest: class imbalance. This chapter teaches you how to visualize it so you can decide whether to resample, reweight, engineer new features, or simply be wiser about metrics.

---

Why this matters (quick refresher)

• Class imbalance biases learning algorithms, evaluation metrics, and even your intuition.
• Visualizations help you see how imbalanced things are, how imbalance interacts with features, and whether minority-class patterns exist or are just noise.
• Builds on pairwise relationships: instead of looking at correlations across the whole dataset, visualize them by class.

Ask yourself: "Does the minority class live in a different part of feature space, or is it completely mixed with the majority?" The answer determines strategy.

---

Core plots and what they tell you

1) Simple class-count bar chart (start here)

Why: the most honest picture of imbalance. Use absolute counts and percentages together.

• What to show: side-by-side bars for counts and an overlaid text with percent.
• Pitfalls: log scales hide absolute scarcity — show both linear and log if counts vary hugely.

Code snippet (seaborn/matplotlib):

import seaborn as sns
import matplotlib.pyplot as plt
sns.countplot(x='target', data=df)
for p in plt.gca().patches:
    plt.gca().annotate(f"{int(p.get_height())}\n({p.get_height()/len(df):.1%})",
                       (p.get_x()+p.get_width()/2, p.get_height()),
                       ha='center', va='bottom')
plt.title('Class counts (absolute and percent)')

2) Class vs. feature distribution (numerical)

• Plot KDEs, histograms, or violin plots by class to see whether the minority class has a different distribution.
• Use transparency and same x-axis limits for fair comparison.

Why: If minority-class density overlaps heavily with majority, resampling alone may not help much — you may need feature engineering.

3) Class vs. feature distribution (categorical)

• Use stacked bar charts showing counts or proportions of categories per class.
• Mosaic plots are great when you want joint-proportions visually.

Important: use proportions (within-class) and absolute counts side-by-side — a rare class may have a strong proportion in a category but still be few in absolute terms.

4) Pairwise plots with stratified sampling

• Pairplots colored by class are excellent but O(n^2) and can get overwhelmed by many points.
• Strategy: subsample majority class to match minority count for visibility, or use alpha and hex/bin plots.
• This is where you build on the previous Pairwise Relationships topic: look at pairwise separation conditioned on class.

5) Dimensionality reduction visualizations (PCA / t-SNE / UMAP)

• Run PCA/t-SNE/UMAP on features and color points by class.
• Use this to explore separability: are minority points clustered or randomly sprinkled?
• Caveats: these techniques can distort distances — don’t over-claim causality.

6) Feature importance / class-conditional feature ranking

• Train a quick tree-based model (with cross-validation) and plot feature importances by how predictive they are for the minority class.
• This is borderline modeling, but it’s useful as a diagnostic during EDA.

7) Correlation and contingency heatmaps per class

• Compute correlation matrices for majority and minority separately and visualize differences.
• For categorical pairs, use Cramér’s V heatmaps by class to see structural differences.

---

Practical examples & metaphors

• Imagine the minority class is a hidden speakeasy in a city: stacked bar charts tell you which neighborhoods (categories) it prefers; KDEs tell you whether it sneaks into similar price brackets as the majority; PCA shows whether it exists in one small cluster or is a bunch of people scattered across the metropolis.

• If the minority is clustered in PCA space, synthetic oversampling (SMOTE variants) might work. If it's scattered and indistinguishable, oversampling could make your model hallucinate.

---

Visualizing sampling strategies (Before / After)

Always visualize the effect of resampling (undersample/oversample/SMOTE) on class counts and feature distributions.

• Plot counts before and after.
• Overlay feature distributions before and after — does SMOTE create unrealistic synthetic examples? If a new synthetic density looks unnaturally smooth or extends into feature regions with zero real points, be suspicious.

Code sketch:

# show distribution before and after SMOTE (example)
sns.kdeplot(data=df, x='feature1', hue='target')
# after resample: df_resampled
sns.kdeplot(data=df_resampled, x='feature1', hue='target', linestyle='--')

---

Pitfalls & how to avoid lying with plots

• Never plot percentages only when class counts differ drastically. A 90% majority and a 10% minority look small in % but can still be large in absolute terms.
• Avoid plotting tiny minority class with same alpha/marker size as majority — it disappears.
• Log scale can be useful but show linear version too so stakeholders understand absolute impact.
• Be careful with overplotting. Hexbin/contour or subsampling for pairplots keeps visual noise down.

---

Quick decision map (visualization → action)

• Minority concentrated in distinct cluster(s): consider oversampling (SMOTE variants), class-weighted loss, or targeted feature engineering for that cluster.
• Minority overlaps heavily with majority: focus on richer features, better feature transforms, or domain-specific signals rather than naive resampling.
• Minority concentrated in certain categories: create interaction features (category x numeric) or target-specific encodings.
• Resampling creates unrealistic feature support: don’t use the synthetic data blindly — consider cost-sensitive learning instead.

---

Checklist: What to plot during your EDA for imbalance

1. Class count bar chart (absolute + percent) — always.
2. KDE/violin/boxplot of top numeric features by class.
3. Stacked bar / mosaic plots for important categorical features.
4. Pairwise scatter (subsampled) or hex/bin plots colored by class.
5. PCA / UMAP / t-SNE colored by class (with caution).
6. Before/After plots for any sampling strategy you might try.
7. Correlation/contingency differences between classes.

---

Closing mic-drop

If you only remember two things:

• Visualize counts in absolute terms, then explore conditional feature distributions by class.
• Use dimensionality reduction and pairwise plots to answer this simple but decisive question: "Do the minority cases live in a different place in feature space, or are they just fewer noisy copies of the majority?"

Do that, and you’ll skip a ton of bad modeling decisions. Go forth, plot ferociously, and never let an imbalanced dataset surprise you at the validation step.