4 ms·
Ask HN: Is trolling on social media a sign that machine learning is overhyped?
YouTube comments, Facebook and Twitter are all full of trolls. *
Facebook also has lots of business profiles (not pages) that aren't trying very hard to pass as people, as well as spam pages, forums and the like also using human profiles
With the rise of organized government trolling ops, this issue is becoming much more important.
You'd think that these leading lights would have figured out how to flag bad accounts more effectively than they do, and to train ML models to find similar accounts to those that are closed after manual reporting. But none of this seems to be the case.
Could it be that they simply lack the ability to do so? And does that tell us anything about the state of AI and Big Data in 2016?
*The classic meaning of "troll" not the generecized insult
- benologist 10y agoIt's a sign of trolling being an unsolved problem, not proof a solution doesn't exist.
- throw_away_777 10y agoMachine learning is a tool that cuts both ways - sites can use it to try and find trolls, but trolls can also use it to better automate bots.
- glondi 10y agoMost of the pages and accounts I refer to are not bots, but businesses with obvious names and content, or are run by real people who are from lower wage countries, and not sophisticated at all.
- auganov 10y agoDo we know for a fact that they want to solve this problem? I'm sure they'd love to change user behavior in a way that would produce more interesting interactions. But I doubt they truly want to go after the random fluff and outright delete it. They'd be penalized for it in the short term (less engagement etc). What really needs to happen is the general public getting fed up with it.
- niftich 10y agoThese are separate issues. From the perspective of a social network operator who is interested in personalized data to collect an accurate social graph and ultimately sell ads (or data), spam accounts are arguably more of a threat than trolls, as spam accounts will never exhibit 'normal human' behavior like providing truthful data about their lives or presenting a truthful network of influence. Troll accounts may in fact promote more engagement (in the most clinical definition of the word), since both trolls and their victims are actual humans. But being harassed drives people away so it isn't just a user retention risk, but also a reputational one. Outside of social networks, a frequent solution to curb harassment is community moderation like downvotes, which only works if the majority is on the side of the victim and not the attacker. When no such consensus exists, all involved parties will be in disagreement with the opposing side, and any administrative to adjudicate a decision on who's in the right will wrong at least half of the participants. ML may be able to filter out blatantly spammy content, but moderating adversarial content is something us humans haven't quite solved either.
- seanwilson 10y agoI'm guessing programmatically identifying trolls is a very hard problem. You could identify the really obvious ones from a blacklist of words (e.g. insulting words, swear words) but I imagine many require real world and topic specific knowledge to identify (e.g. implying silly things are true, insults that require parsing a sentence or two to understand).