6 ms·

Indeed. The title is misleading. From the PDF: "In the two months preceding the election, around 57,000 users posted approximately 426,000 unique Tweets contai

Indeed. The title is misleading. From the PDF: "In the two months preceding the election, around 57,000 users posted approximately 426,000 unique Tweets containing variations of the #PodestaEmails hashtag. Approximately one quarter (25%) of those Tweets received internal tags from our automation detection systems that hid them from searches. As described in greater detail below, our systems detected and hid just under half (48%) of the Tweets relating to variants of another notable hashtag, #DNCLeak, which concerned the disclosure of leaked emails from the Democratic National Committee. These steps were part of our general efforts at the time to fight automation and spam on our platform across all areas."

creaghpatr 9y ago
I'm interested to know how they determined half of all #DNCleaks tweets were spam and how that compares to other trending topics both political and apolitical. Seems like they are hiding behind the anti-spam stance without detailing how their system determines that. Removing almost 50% of communication on a topic is a pretty dramatic step, especially during an election.
- teraflop 9y ago
  Yes, but it would be an equally dramatic step to leave their normal anti-spam system in place, but then specifically exempt a particular hashtag from filtering. But of course you're right; the core problem is that the filtering process itself is completely opaque and inscrutable. We just have to take Twitter's word that it doesn't introduce explicit or implicit biases.
- spdustin 9y ago
  In my experience analyzing these sorts of things for my skunkworks project, the "source" (client or app) and too-quick-to-be-human retweet brigades are dead giveaways, and I'd imagine they're part of this algorithm. I also started tracking "visual hashes" of avatars (blurred and reduced to a 32x32 grid) in an effort to find re-used or stock profile pics, which has shown some promise as a spammer/bot/brigade signal. As well, existing known signals such as "user with default background color" or "no verified email address" could be part of it. For example, IFTTT was recently shown to be used by a pro-Trump "retweet brigade" attempting to boost the "Uranium One" campaign with old "RT-style" retweets that look like new tweets but all have the same text. That group wasn't terribly smart, since they all looked the same, and were broadcast in relatively short order by hundreds of accounts. I know these are guesses; Twitter likely won't ever share their secret sauce recipe.

creaghpatr 9y ago

I'm interested to know how they determined half of all #DNCleaks tweets were spam and how that compares to other trending topics both political and apolitical. Seems like they are hiding behind the anti-spam stance without detailing how their system determines that. Removing almost 50% of communication on a topic is a pretty dramatic step, especially during an election.

teraflop 9y ago
Yes, but it would be an equally dramatic step to leave their normal anti-spam system in place, but then specifically exempt a particular hashtag from filtering. But of course you're right; the core problem is that the filtering process itself is completely opaque and inscrutable. We just have to take Twitter's word that it doesn't introduce explicit or implicit biases.
spdustin 9y ago
In my experience analyzing these sorts of things for my skunkworks project, the "source" (client or app) and too-quick-to-be-human retweet brigades are dead giveaways, and I'd imagine they're part of this algorithm. I also started tracking "visual hashes" of avatars (blurred and reduced to a 32x32 grid) in an effort to find re-used or stock profile pics, which has shown some promise as a spammer/bot/brigade signal. As well, existing known signals such as "user with default background color" or "no verified email address" could be part of it. For example, IFTTT was recently shown to be used by a pro-Trump "retweet brigade" attempting to boost the "Uranium One" campaign with old "RT-style" retweets that look like new tweets but all have the same text. That group wasn't terribly smart, since they all looked the same, and were broadcast in relatively short order by hundreds of accounts. I know these are guesses; Twitter likely won't ever share their secret sauce recipe.