Thread by @svpino

Two helpful metrics to evaluate a machine learning model: Sensitivity and Specificity.

Here is how they work: ↓

2. I'm mainly used to thinking about Precision and Recall, but these new metrics come in handy when working with a ROC curve.

They are less popular in the machine learning community but widely used in other fields.

3. Let's start with Sensitivity.

Sensitivity → True Positive Rate. The capacity of a model to identify positive samples.

Sensitivity = (TP) / TP + FN

This should look very familiar: Sensitivity and Recall are the same things!

4. How about Specificity?

Specificity → True Negative Rate. The capacity of a model to identify negative samples.

Specificity = (TN) / TN + FP

Warning: Specificity is not the same as Precision!

5. Let's use an example and use all three metrics: Precision, Recall, and Specificity.

(Remember, Sensitivity and Recall are the same, so we don't need to double down.)

6. Imagine a model trained to classify patients sick with Covid 19.

• Precision: Among all of the patients that the model classified as being sick, how many were ill?

• Recall: How many did the model correctly classify among all the sick patients?

7. And finally,

• Specificity: Among all healthy patients, how many did the model correctly diagnose as being healthy?

8. To recap:

• Sensitivity → True Positive Rate of the model.

• Specificity → True Negative Rate of the model.

• Higher sensitivity → Model is good at detecting positive samples.

• Higher specificity → Model is good at detecting negative samples.