Multicategory Choice Modeling by Recurrent Neural Nets
(A2024-118711)

AUTHORS

Harald Hruschka, University of Regensburg

ABSTRACT

We investigate three variants of recurrent neural nets capable to reproduce dynamic effects in a flexible way. We compare these recurrent nets to non-recurrent multilayer perceptrons (MLPs) and to multivariate logit models. The latter are often used to analyze multicategory choice. Models are evaluated by binary cross-entropies for a holdout sample of households. A six hidden variable MLP turns out as best non-recurrent model. Overall, a recurrent long-short term memory net with six hidden variables outperforms the other models considered. We restrict further analyses to these two neural nets, which both include category-specific features as input variables. We compute average marginal effects of category-specific features. An evolutionary algorithm serves to determine optimal weekly category-specific features. Optimization based on the recurrent net provides an average revenue per basket higher than its observed value. As a rule, optimal category-specific features change considerably across weeks.