Module 07 Summary

What this module established

Accuracy on imbalanced data reports the class balance. Report recall on the class you care about, alongside the majority baseline - a useless constant predictor can beat a model that finds half the damage.

Carry forward

• Read the confusion matrix with the cost of each error attached. The ratio between them determines the operating point.
• Choose the threshold by minimising error cost. It is usually the cheapest available improvement, and 0.5 is rarely right.
• Sort errors by confidence: a confident mistake means a learned rule is wrong, a borderline one means the case was hard.

Before moving on

Move on when your threshold comes from a cost ratio and your slices were defined before you saw any results.

Mark complete