My questions is as above. What are the most important matrices (f1, precision, recall...etc) that I need to prioritize my work to improve for evaluating how good a model predict customer churn and the reason behind it.
For example: a fraud detection data product that had a business constraint: zero false positives. I set the constraint myself.
The reason for it is that we ban users based on that data product’s output. It is absolutely inadmissible that a user is banned incorrectly based on the output of an algorithm.
In this specific case I prefer to have dozens, even hundreds of false negatives than one single false positive.