Skip to content

Class Imbalance Handling

Strategies for dealing with imbalanced classes in the Titanic dataset.

Problem

The Titanic dataset has imbalanced classes: - 62% Did Not Survive (0) - 38% Survived (1)

This can lead to models biased toward the majority class.

Solution: Stratified Cross-Validation

All cross-validation uses StratifiedKFold to maintain class distribution in each fold.

from sklearn.model_selection import StratifiedKFold

cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)

Benefits

  • Maintains exact class proportions in each fold
  • More reliable performance estimates
  • Reduces variance across folds
  • Better for small datasets

Additional Techniques

1. Class Weights

Penalize misclassifications of minority class more heavily:

from sklearn.linear_model import LogisticRegression

model = LogisticRegression(class_weight='balanced')

2. SMOTE

Synthetic Minority Over-sampling Technique:

from imblearn.over_sampling import SMOTE

smote = SMOTE(random_state=42)
X_resampled, y_resampled = smote.fit_resample(X_train, y_train)

3. Ensemble Methods

Use models robust to imbalance: - Random Forest: Built-in class weighting - XGBoost: scale_pos_weight parameter - CatBoost: Handles imbalance automatically

4. Evaluation Metrics

Use metrics appropriate for imbalanced data: - F1-Score: Harmonic mean of precision and recall - ROC-AUC: Threshold-independent - Precision-Recall Curve: Better than ROC for imbalance

Current Implementation

The project uses:

  1. ✅ Stratified K-Fold CV (5 folds)
  2. ✅ Weighted F1-Score as secondary metric
  3. ✅ XGBoost/CatBoost (imbalance-aware models)
  4. ✅ ROC curves for threshold analysis

When to Use Each Technique

Technique Use When Avoid When
Stratified CV Always (best practice) Never
Class Weights Moderate imbalance (30-70%) Extreme imbalance (<5%)
SMOTE Severe imbalance (<20%) Sufficient minority samples
Ensemble Any imbalance Need interpretability

Further Reading