• Table of Contents

Basic Classification Metrics

Although this is about classification, we can equivalently, and more intuitively frame things in terms of selection. Imagine yourself as a fisher hauling up a catch:

• “positives” = your net
• “negatives” = everything outside your net
• Precision = “how much of the things caught are fish”?
• Recall = “how much of the fish out there did we actually get?”

→ The most common metrics as sentences you’d probably see in real life:

1. Precision: How pure is our selection?
2. Recall: Did we get them all?

and some of the slightly less intuitive interpretations:

• NPV: how pure is the anti-selection?
• Specificity: did we leave them all?

Accuracy

More generally, notice that any confusion matrix reduces to a TP/FP/TN/FN matrix in a global, or relative (wrt one class) context.

• accuracy = Measuring the diagonal

Precision