What is an imbalanced dataset?

An imbalanced dataset has classes with very different frequencies.

Example: fraud detection

• 99.5% normal
• 0.5% fraud

If you predict “normal” always, you get 99.5% accuracy… but the model is useless.

Why accuracy becomes misleading

With imbalance, you should prioritize:

• precision
• recall
• F1-score
• PR-AUC

Strategies to handle imbalance

1) Use better metrics

Always start here.

2) Use class weights

Many models support weighting the minority class more.

Example: Logistic Regression

• class_weight="balanced"class_weight="balanced"

3) Resampling

• Undersampling majority class (risk: lose information)
• Oversampling minority class (risk: overfit)

4) SMOTE (Synthetic Minority Over-sampling Technique)

SMOTE creates synthetic minority samples by interpolating between existing ones.

  flowchart LR
  A[Minority sample i] --> C[Synthetic sample]
  B[Nearest minority neighbor] --> C

Important warning: avoid leakage with SMOTE

SMOTE must be applied only to the training set.

Best practice:

• apply SMOTE inside a cross-validation pipeline

Practical note

SMOTE is commonly used via imblearnimblearn (imbalanced-learn library). If you don’t want extra dependencies, start with class weights.

Mini-checkpoint

Given a dataset with 95/5 split:

• which metric would you report?
• what might be the business cost of false negatives vs false positives?