Comment
Comment by londons_explore
The training data doesn't seem to be great quality...
I think better results could be obtained by manually going through these 52,000 training examples - even a couple of seconds per example should be enough to weed out obviously wrong ones, and should only take <$1000 of human time.