A false positive is a negative case predicted as positive, like a legitimate email sent to spam. A false negative is a positive case predicted as negative, like a missed tumor. In hypothesis testing they are Type I and Type II errors.
Card 1 of 270. Answer: A false positive is a negative case predicted as positive, like a legitimate email sent to spam. A false negative is a positive case predicted as negative, like a missed tumor. In hypothesis testing they are Type I and Type II errors.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
A false positive is a case whose true label is negative but which the model predicts as positive: a legitimate email sent to spam, or a smoke alarm set off by toast. A false negative is the reverse, a real positive the model calls negative: a tumor the scan misses, or spam that reaches the inbox. They are the FP and FN cells of the confusion matrix. In hypothesis testing the matching errors are a Type I error, rejecting a true null hypothesis, and a Type II error, missing a real effect.
Each has its own rate, measured against a different group. The false positive rate is FP / (FP + TN) = 1 − specificity, the share of actual negatives wrongly flagged. The false negative rate is FN / (TP + FN) = 1 − recall, the share of actual positives missed. Of 5,000 circuit boards, say 80 are faulty and the model flags 68 of them plus 20 good ones. The 12 faulty boards that ship give a false negative rate of 12 / 80 = 15%, yet accuracy is 99.4% because good boards dominate.
A tiny false positive rate can still mean most alarms are false. A moderation model reviews 50,000 harmless posts and 200 abusive ones, removing 250 harmless and 150 abusive posts. The false positive rate is 250 / 50,000 = 0.5%, yet 250 of the 400 removals, 62.5%, were mistakes. That 62.5% is the false discovery rate, FP / (FP + TP) = 1 − precision, and confusing it with the false positive rate is a classic interview trap.
The two errors trade off through the decision threshold: raising it cuts false positives and adds false negatives, and lowering it does the reverse. Which matters more depends on what each costs. When a miss is expensive, as in cancer screening or safety defects, lower the threshold and accept more false alarms; when a false alarm is expensive, as in blocking legitimate payments, raise it. With calibrated probabilities, predict positive when p > C_FP / (C_FP + C_FN): if a miss costs 9 times a false alarm, the threshold drops from 0.5 to 0.1.
