Hacker News Reader

Comment by lazyasciiart | Hacker News Reader

Not quite true. A high false positive rate leads the people running the machines to stop reacting seriously to alerts, which means that they don’t do the manual checks with as much care. This is even more of an issue when everyone in line knows they are bullshit and has no respect for the people running them.

View on HN

A false positive rate of close to 50% means that every second alert indeed pointed at something real. That should not cause the staff to get unmotivated. Again, without the false negative rate we can't judge if they are complete bullshit or not. If it were a 90% false positive rate then sure, that would clearly have the effects you outlined. I don't think at around 50% that's the case yet.

> A false positive rate of close to 50% means that every second alert indeed pointed at something real.

No it doesn't.

Please see my other comment which references the definition of false positive rate because I think your understanding of the term is wrong. You have a 50% alarm/trigger rate in mind. And please try to reply with a bit more substance than "No".

To quote an article on the subject: "meaning that every other person who went through the scanner had to undergo at least a limited pat-down that found nothing"

> And please try to reply with a bit more substance than "No".

Eh. I was hoping you'd just delete the comment rather than have two parallel arguments.

> The false positive rate is calculated as the ratio between the number of negative events wrongly categorized as positive (false positives) and the total number of actual negative events (regardless of classification).

https://en.m.wikipedia.org/wiki/False_positive_rate

54% FPR means half of all people who aren't trying to bring anything illegal onto the plane gets stopped.

Since basically nobody are actually trying to do this, that also means that more than half of all people going through gets stopped.

The ratio of true positives to positives must thus also be vanishingly small.

No, it means that about 50% of people scanned would trigger a "positive". The false positive rate is the number of false positives divided by the number of actual negatives.

100 people, 50 false alarms, 1 actual problem alerted and 49 marked as safe correctly = 50/(50+49) = ~50%.

That's just wrong. This is ML 101

False positive = (Machine says positive but grand truth is negative) / (Machine says is positive)

In English, a false positive is when the machine declares something as positive but it is actually negative.

This fits a sanity test because ~50% of people that pass through a detector do not get patted down. The truth is closer to 10%. Even then, a true positive probably includes someone with two pennies in their pocket.

https://en.m.wikipedia.org/wiki/False_positive_rate

In statistics and ML it’s (FP) / (FP + TN) aka (false positives) / (actual negatives samples). This is by far the most common definition.

Here ML using the common definition see fall-out: https://en.m.wikipedia.org/wiki/Precision_and_recall

You are right this is ML 101, but you are wrong. Someone else has linked the wiki page, and I've posted it elsewhere and I would strongly recommend you read it as your interpretation is very incorrect.

> In English, a false positive is when the machine declares something as positive but it is actually negative.

Yes, and the rate is based on the frequency this happens in your negative set.

>This fits a sanity test because ~50% of people that pass through a detector do not get patted down.

The 50% figure may be wrong, but my definition of what a false positive rate is not.

You're thinking of FDR: False discovery rate. It equals Σ False positive/Σ Predicted condition positive.

See https://en.m.wikipedia.org/wiki/Precision_and_recall for a table of different terms.

Wow, that definition is counter-intuitive. Ive expected the false-positive-rate to be the number of false positives divided by the number of all positives. Isn't it how "rates" are usually defined? No wonder everybody is confused.

I think it's a perspective thing. You are looking at the rate of false results, so given a particular type of input what rate do you get it right? For negative ground truth your possible outputs are "false (positive)" and "true (negative)".

I agree it can be confusing, but I think people would also make the opposite mistake if it were defined differently.

You're thinking of FDR: False discovery rate. It equals Σ False positive/Σ Predicted condition positive.

See https://en.m.wikipedia.org/wiki/Precision_and_recall for a table of different terms.

Reply on news.ycombinator.com