When would you optimize for recall over precision, and how do you actually set the threshold?

Optimize recall when a missed positive costs more than a false alarm, such as fraud screening or disease detection where a human reviews every flag anyway. Optimize precision when acting on a positive is expensive or annoying, such as sending offers or blocking accounts. Set the threshold from expected cost rather than defaulting to 0.5, by sweeping thresholds and picking the minimum expected loss.

I start by asking what happens to each kind of mistake. On a fraud queue, a false negative is money out the door plus a chargeback, while a false positive is thirty seconds of an analyst's time, so recall wins and I push the threshold down. On a push notification model it flips: a false positive irritates a real user and can cost you an uninstall, so I want precision even if I reach fewer people. To pick the actual number, I put a rough cost on each error type, sweep the threshold across the validation set, and plot expected cost. It usually has a clear basin, and I take the middle of the basin rather than the exact minimum, because the minimum jumps around between folds. The constraint people forget is capacity. On one fraud project the cost curve said flag about four percent of transactions, but the review team could handle roughly a thousand cases a day, so the real threshold was whatever produced a thousand. I said that plainly rather than pretending the choice was purely statistical.