Data scientist interview: precision vs recall — how do you choose which to optimise?

Model answer

Define both plainly, then make the point that the right choice depends on the cost of each error. Precision asks: of the cases the model flagged as positive, how many really were? Recall asks: of all the actual positives, how many did the model catch? There is usually a trade-off — pushing one up tends to pull the other down as you move the decision threshold. The key is to reason from consequences. When a false positive is costly — say flagging a good customer as fraud and blocking them — you favour precision. When a false negative is costly — say missing a disease or a real fraud — you favour recall. Give a concrete example for the problem at hand and state which error hurts more. Mention that when you need a single balanced measure you can use the F1 score, and that you would look at the precision-recall curve and choose the threshold with the business, not in isolation. Avoid claiming one is always better. Ending with "so I would ask what a mistake costs us in each direction, then set the threshold accordingly" shows real judgement.

This is a guide to structure and approach, not a script to memorise — say it in your own words.

Practice this yourself

Type your own answer below and get it scored instantly by the same AI that powers Vachix mock interviews.