Capturing Heterogeneity Among Consumers with Multi-taste Preferences

Previous research has suggested that an individual consumer may have multiple tastes within a given product category. Multi-taste preferences are likely present in categories characterized by large product attribute spaces and many diverse products, such as music, videos, restaurants, or books. Capturing heterogeneity among such multi-taste consumers requires new methods, as two consumers may share some tastes but not others. This is a different type of heterogeneity than that captured by existing models, such as mixed logit or latent class models, which estimate only one taste for each individual. In this paper, we propose a model that allows for heterogeneity among consumers with multiple tastes, and an estimation procedure that scales to potentially very high dimensional attribute spaces. In a numerical study, we simulate consumers with multiple preferences and demonstrate the proposed algorithm accurately recovers parameters, whereas single-taste benchmark models underfit and lead to a misleading picture of individual level preferences and the population preference distribution. We then test the algorithm empirically on a data set of recipe choices. We show that the algorithm scales to a large parameter space, and that the model fits the data better than single-taste benchmarks. We also demonstrate that the model uncovers rich patterns of underlying heterogeneity, such as what the different tastes are, how many consumers have each taste, and which tastes tend to be more and less likely to occur in the same individual.