Sequential-Set Recommendation

Our methods so far ignore time which is an important variable that can be monitored in almost any application. In this chapter, we extend item recommendation (see chapter 6) with time information. In general, time is a continuous variable with infinite support. Thus, our factorization models in chapter 5 cannot be applied directly as they assume a categorical domain. Also simple discretization of the domain would not work because (1) factorization models assume no a priori relationship between two variable instances (e.g. two close points in time) and (2) the model could not predict in the future as no observations for these variables are present. Thus, our approach is different: we reformulate the problem with sequences and use the independence assumptions of Markov chains within our model. That means for each user, we see his action of the past as a sequence – e.g. what products he has bought. Typically, several products are bought at the same day and thus, we have per user a sequences of sets (=baskets/ shopping carts). The Markov chain assumption is now that the next action (=shopping cart) of the user depends only on a few of his previous ones.