I've experimented with classifying my own transaction history using n-grams and either naive Bayes or a neural network, the neural network worked better. I could train the nn to convergence in a few seconds and it was good enough to remove 95% of the classification work.