I don't buy it. They just found a correlation that users with a downvoted comment happened to also make other bad comments (what a surprise.) They didn't actual

I don't buy it. They just found a correlation that users with a downvoted comment happened to also make other bad comments (what a surprise.) They didn't actually downvote or upvote comments on randomly picked users and see what effect it had, which would have been way easier than the machine learning approach they used.