Precision vs Recall | Machine Learning

Precision and Recall are two very important metrics for a Classifier ML Model. One easy way to remember the definitions is:

Precision

What the classifier is saying is infact true

Let’s say your classifier has a 90% precision. It means your classifier keeps its word. When it says it sees an apple, you can consider thats 90% true and with 10% to chance.

Recall

How much the classifier identifies among all the true instances.

Let’s say your classifier still got 90% precision but at a recall of 30%. This recall of 30% means that say among all the apples, your classifier identified (detected) only 30% of the apples. The rest 70% of apples was never found by your classifier even though its precision is at 90% keeping its word.

Maximize the Trade Off

There are times when you want to make use of this trade off between Precision and Recall and use it in your application.

For instance, in the example of Apple vs Orange sorting machine, you want your Precision to be really high if your machine is going to be identifying and packing Apples. As a customer, you never want to see an Orange in the box of Apples you just bought at Walmart, right? In this case, 99% Precision and a low recall could be fine for your Classifier. To manage with the low recall, the machine would have to have some human help to pick out the left out Apples from the batch that was already processed.

A High Recall classifier might be useful in domains where you never want the event to occur at any cost. For example, in a Classifier deployed in a Banking domain that tries to prevent Cheque Fraud, you want a high Recall %, meaning, the Classifier should identify all (or atleast try to maximize) the instances of Fraudulent signatures. It doesn’t matter if the classifier identifies it as Fraud but later a teller verifies it to be legitimate. (Thats what the Bank Employees are for, right?!). In this scenario, a high recall is needed instead of a high precision.

Memes

Finally, I had great fun going through all the memes on Google on this subject. Memes are an awesome way to relate and “burn” these ML concepts in your brain forever! I also ended up creating my own, (the Thanos, at what cost? is mine!)