As humans, we've always made decisions using these kinds of metrics — even if we didn't know their names or realize it. Let me help you in realizing it.
Machine Learning Algorithms can be broadly categorized into supervised, unsupervised, and reinforcement learning. In this article we focus on classification model metrics.
Police hired a private investigator to identify murderers among innocents. The investigator has 7 days. On which day did he perform best?
Use your gut — don't think about metrics yet.
The answer depends entirely on the context and the problem you're trying to solve. Once the problem is well-defined, the right metric becomes obvious.
"It is better that ten guilty persons escape than one innocent suffer."
Minimize false positives. When the system flags someone as guilty, it must be very likely correct.
Drones scanning a forest after a plane crash. The goal is to find every single person who might be alive. Missing someone could cost a life.
False alarms are acceptable — rescuers can investigate. But missing a real survivor? Never.
High Precision, High Recall, and High Accuracy is ideal — but far from reality. Just like infinity, we can always get closer and closer, but never reach 100%.
It therefore becomes essential to ask your business: which metric makes you most profitable, and which can you slightly compromise on — based on your core problem.