Bayesian Modeling of Consumer Behavior in the Presence of Anonymous Visits

Tailoring content to consumers has become a hallmark of marketing and digital media, particularly as it has become easier to identify customers across usage or purchase occasions. However, across a wide variety of contexts, companies find that customers do not consistently identify themselves, leaving a substantial fraction of anonymous visits. We develop a Bayesian hierarchical model that allows us to probabilistically assign anonymous sessions to users. These probabilistic assignments take into account a customer's demographic information, frequency of visitation, activities taken when visiting, and times of arrival. We present two studies, one with synthetic and one with real data, where we demonstrate improved performance over two popular practices (nearest-neighbor matching and deleting the anonymous visits) due to increased efficiency and reduced bias driven by the non-ignorability of which types of events are more likely to be anonymous. Using our proposed model, we avoid potential bias in understanding the effect of a firm's marketing on its customers, improve inference about the total number of customers in the dataset, and provide more precise targeted marketing to both previously observed and unobserved customers.