> If Twitter knew how to count all the spam on the platform, it would know how to eliminate all the spam on the platform.
Uh. Not at all. You can give an accurate count of something through costly sampling, but the same mechanism can't be used to eliminate it because it would be uneconomical to sample more than the small amounts required to give an accurate figure. This is doubly so when even a successfully identified spammer can just come back under another identity-- identifying the spammers isn't enough to eliminate them, even if you can afford to do so.
Otherwise-- it all sounds reasonable, and I've been enlightened a bit on the whole 5% debacle.
Sampling is highly effective. With a large population for most distributions it takes only a fairly small sample before the sampling noise falls under the process noise: the number of active spammers isn't a constant, so even if you could count every one of them the number would be constantly changing, there is usually little point in knowing the count with more accuracy than the random/shot-noise component of that time-to-time variation. In so far as there could be said to exist a number at all, sampling will usually get you there when dealing with a large population. And often much better, because with sampling you can often put in place processes to better control bias sources than you could in an effort to just count everything (e.g. by better characterizing your samples).
When it comes to addressing the problem instead of measuring it, hitting a small percentage of it would have an insignificant effect. So it's just not correct to say that being able to measure the proportion spam with high precision would automatically mean that they were able to solve it.
Nor is it necessarily the case that they'd want to-- the argument that they're incentivized to stop it from being "too great" doesn't mean that the spam isn't irrelevant to or even beneficial to the company or its employees at lower levels than that. Particularly to the extent that the tolerable level of too greatness depends on the degree to which they can conceal the true level of spammers from advertisers and investors.
Imagine that the MAU figure is calculated in part by tracking clickthroughs and then some spammers figure out that they can reduce their odds of being banned by making their spambots click ads. Suddenly spam (or that portion of it, at least) is in twitters interest up until the point that it visibly hurts conversion rates and considering that conversion is a much smaller and noisy figure and that many campaigns can't track conversion at all... But only so long as they can maintain plausible blindness to this activity and keep it secret.
Genuine thank you for the lesson for myself and the rest of us, but holy fuck bud, that's a lot of words to avoid saying it's a sample-based estimate, not a count!
5 comments
[ 4.1 ms ] story [ 20.7 ms ] threadUh. Not at all. You can give an accurate count of something through costly sampling, but the same mechanism can't be used to eliminate it because it would be uneconomical to sample more than the small amounts required to give an accurate figure. This is doubly so when even a successfully identified spammer can just come back under another identity-- identifying the spammers isn't enough to eliminate them, even if you can afford to do so.
Otherwise-- it all sounds reasonable, and I've been enlightened a bit on the whole 5% debacle.
When it comes to addressing the problem instead of measuring it, hitting a small percentage of it would have an insignificant effect. So it's just not correct to say that being able to measure the proportion spam with high precision would automatically mean that they were able to solve it.
Nor is it necessarily the case that they'd want to-- the argument that they're incentivized to stop it from being "too great" doesn't mean that the spam isn't irrelevant to or even beneficial to the company or its employees at lower levels than that. Particularly to the extent that the tolerable level of too greatness depends on the degree to which they can conceal the true level of spammers from advertisers and investors.
Imagine that the MAU figure is calculated in part by tracking clickthroughs and then some spammers figure out that they can reduce their odds of being banned by making their spambots click ads. Suddenly spam (or that portion of it, at least) is in twitters interest up until the point that it visibly hurts conversion rates and considering that conversion is a much smaller and noisy figure and that many campaigns can't track conversion at all... But only so long as they can maintain plausible blindness to this activity and keep it secret.