There are two big problems that we deal with. First is estimation of customer lifetime value. We use a latent attrition model, which is the 'pattern matching'

The second is figuring out which promotions/emails go to which people. This is a supervised learning problem. We train the model with users past responses to discounts and their past behavioral states (which are the posterior probabilities from the latent attrition model). Then we use this to predict how users in those states will respond to similar promotions in the future.