Statistics for ML #78 — Imbalanced Data — SMOTE & Class Weights
Published:
Imbalanced Data — SMOTE & Class Weights
| Post #78/100 in the Statistics for ML series — Md Salek Miah | Statistician & ML Researcher | SUST, Bangladesh. |
Imbalanced data is ubiquitous in health outcomes research. In DHS data, rare outcomes like maternal mortality or specific diseases can have extreme class imbalance.
Why Standard Models Fail
A model that always predicts the majority class achieves high accuracy but is clinically useless.
Example: If only 5% of women experienced pregnancy loss, a trivial model predicting “no loss” for everyone achieves 95% accuracy but 0% sensitivity.
Solutions
| Method | Description | When to use |
|---|---|---|
| SMOTE | Synthetic Minority Over-sampling | Small-medium datasets |
| Class weights | Penalise majority misclassification | Any size, simpler |
| Threshold tuning | Adjust decision threshold | Post-training |
| Ensemble methods | BalancedBaggingClassifier | Works well in practice |
| Precision-Recall AUC | Better metric than ROC-AUC | Severe imbalance |
from imblearn.over_sampling import SMOTE
from imblearn.pipeline import Pipeline
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import classification_report
# SMOTE + Random Forest pipeline (used in my DHS research)
pipeline = Pipeline([
('smote', SMOTE(random_state=42, sampling_strategy=0.5)),
('clf', RandomForestClassifier(class_weight='balanced', n_estimators=200))
])
pipeline.fit(X_train, y_train)
print(classification_report(y_test, pipeline.predict(X_test)))
Series Index | Post #78/100 | Md Salek Miah | saleksta@gmail.com