False positives & negatives, explained simply
A yes-or-no model can be wrong in two different ways. A false positive is a false alarm: it says yes when the answer is no. A false negative is a miss: it says no when the answer is yes.
Many models give a score and use a threshold to decide. Lower the threshold and the model catches more real cases but raises more false alarms. Raise it and false alarms drop, but more real cases slip through.
Which mistake is worse depends on the job. For a smoke alarm, a miss is far more dangerous than a false alarm. For a spam filter, sending a real email to spam can be the bigger problem. People have to decide, not the model.
Example: A mountain rescue alarm
- The alarm listens for calls for help.
- False positive: it rings for a bird call. Annoying, but everyone is safe.
- False negative: it stays silent when a hiker shouts. Dangerous.
- So rescuers set it to be extra sensitive and accept more false alarms.
Try this
- For a school fire alarm, a spam filter and a "is this a cat photo" app, decide which mistake is worse.
- Play a listening game: raise your hand at the sound of a word. Be very strict, then very relaxed. Count both kinds of mistake.
- Explain to a grown-up why a model with fewer total mistakes is not always the better one.