Revision

Back to ML System Design

Introduction

Different solutions exist to combat class unbalanced.

Right metric

Choose the right metrics: F1 Score, ROC curve, AUC, etc. Generally an asymetric metric is needed for unbalanced data.

Data-level methods

See the page on Resampling that talk about Tomek Links and SMOTE.

Algorithm-level methods

See the page on Algorithm-level methods that talk about Cost-sensitive learning, Class-balanced loss and Focal loss.

Resources

See:

• CS329 lecture 4.