The data was a huge amount of complete url history and quite a lot of feature engineering goes into developing the features. Mix of Gaussian processes, random forests, GDBT, and a lot of hand coded stuff using regexes and state machines. To get the embedding the features were fed into a multi layered RBM auto encoder. Despite quick initial success of the embedding I was unable to evangelize deep learning to the org and left out of frustration. Andrew Ng, before he was famous, visited us to evangelize deep learning and my colleagues just berated him with their views about how much better Bayesian is, it was embarrassing.