Not convinced their user embedding creation is useful. Did not read in detail but it seems to use a list of edits similar to how one may create paragraph vectors as an average of word vectors. But if I had to guess, they're not really capturing more information than they originally had with a one hot vector of whether or not a user had edited a specific article. It would have been better if they had bench marked against this. I would wager that a simple random forest and the one hot vector would do just as well if not better than their NN solution.