This is somewhat interesting. If what is stated in the comment itself is true then that means the poster had to solve to find x and y such that 'This comment is going to collect the most votes yet is predicted to be Bad: x - Good: y' would give Bad: x and Good: y as the output when passed through the neural net. Maybe they also manipulated other parts of the comment to find x and y.
I didn’t read the code in the post and don’t have any deep familiarity with machine learning, but I have implemented a naive bayesian classifier to do something similar for tweets. The scores you get from that method don’t add up to 1 either.
