CS5720 - Handling Imbalanced Datasets

The Imbalance Problem

Class imbalance occurs when one class significantly outnumbers others in your dataset, leading to biased models that ignore minority classes.

Why it's a problem:

• Models learn to predict the majority class
• High accuracy can be misleading
• Minority classes get ignored
• Real-world consequences can be severe

🚨 Real Example

Medical diagnosis: 95% healthy patients, 5% with disease. A model predicting "always healthy" gets 95% accuracy but misses all sick patients!

Solution Strategies

⚖️

Sampling Techniques

Oversample minorities or undersample majorities
🔬

Data Synthesis

Generate synthetic examples using SMOTE, GANs
🎯

Loss Modification

Weighted loss, focal loss, cost-sensitive learning
📊

Better Metrics

Use precision, recall, F1-score, AUC instead of accuracy

Interactive: Before vs After Balancing

Imbalanced Dataset

Class A
(Minority)

950

Class B
(Majority)

Ratio: 5% vs 95%

Balanced Dataset

500

Class A
(Balanced)

500

Class B
(Balanced)

Ratio: 50% vs 50%

Accuracy

95%

Can be misleading!

Precision

12%

Shows true performance

Recall

Captures minority class

Handling Imbalanced Datasets

The Imbalance Problem

Solution Strategies

Interactive: Before vs After Balancing

Modal Title