3 ms·
That says a little bit more. Do you treat it as a reinforcement learning (RL) problem, or as a classification problem? It seems like a sequential decision makin
by noelwelsh 15y ago
That says a little bit more. Do you treat it as a reinforcement learning (RL) problem, or as a classification problem? It seems like a sequential decision making problem, so RL is appropriate but AFAIK there is no RL algorithm that generalises over states while still retaining some error bounds. I suppose you could brute-force a Bayesian solution via MCMC.
- aaronjg 15y agoThere 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.