Multicategory choice modeling by recurrent neural nets

In multicategory choice, a customer may purchase multiple products or product categories at the same time. Hidden variables of recurrent nets depend on current inputs and hidden variables of the previous period. We investigate the three main variants of recurrent neural nets, which we compare to multilayer perceptrons and multivariate logit models. Model evaluation is based on binary cross-entropies for a holdout sample. We restrict further analyses to the best non-recurrent model, a multilayer perceptron, and the best performing recurrent neural net, which both include category-specific advertising (features) as inputs. We interpret these two models looking at category dependences and feature effects. Category dependences measure the strength of either complementary or substitutive relations. We show what the stronger dependences inferred from the recurrent net imply for cross-selling decisions. We also compare what these two models imply for sales promotion by optimizing features. For the multilayer perceptron we obtain features for each category, which are constant across weeks, equaling either zero or the maximum value. For the recurrent net, features assume many intermediate values and vary considerably across weeks. To illustrate managerial implications of the recurrent net, we determine weekly features for six selected categories that differ as much as possible from each other. Finally, we discuss limitations of our approach and opportunities for future research.