zlacker

>>pg+(OP)
"The problem has several components: comments that are (a) mean and/or (b) dumb that (c) get massively upvoted."

Find a few examples of comments that are unambiguously (a), (b), and (c) and have either you personally or someone you trust flag them as such. Next, take the set of all people who upvoted the abc-flagged comments. Their votes now have a 50% chance of not counting towards vote totals from now on, but in a way that the user isn't shown that their votes aren't being counted -- perhaps with an artificial "offset" vote that appears a few minutes later.

There's fun parameters one could throw in there too, like exponential decay on the likelihood of a vote being magically offset that spikes back up every time the user votes stupidly.

>>goodsi+P2
Unfortunately I know that won't work because I've already tried that experiment. For about a year I've been annotating such comments, but when I analyzed who upvoted them, the upvotes were very broadly distributed.

>>pg+r3
Would it be possible to make public a data set of heavily up-voted comments that suck and challenge the community to come up with algorithms that can predict whether a given popular comment sucks or not? I'd think you could get a fair amount of mileage just by measuring frequencies of certain technical noun phrases.

I paused for a second at the privacy implications of this, but honestly all comments are submitted here with the assumption that they will forever be visible and subject to moderation or critique.