Essays on Bayesian Machine Learning in Marketing

 Chapter 1, titled “Privacy Preserving Data Fusion”, and joint work with Longxiu Tian and Dana Turjeman, tackles the complex problem of merging multiple datasets while ensuring user privacy. This paper introduces a privacy-preserving data fusion methodology that adheres to the principles of differential privacy, leveraging variational autoencoders and normalizing flows to create a robust, nonparametric, Bayesian generative modeling framework. This methodology notably accounts for missingness in each dataset, correcting for sample selection and negating the requirement for identical users across datasets when learning the joint data generating process. Through a series of simulations and an applied case involving a novel large-scale customer satisfaction survey and CRM database from a leading U.S. telecom carrier, we demonstrate the potential of this privacy-preserving methodology for robust data fusion, providing insights into customer satisfaction and churn propensity without compromising privacy. 

Chapter 2, titled “Understanding Consumer Expenditure Through Gaussian Process Choice Models”, is joint work with Alan Montgomery. This chapter challenges the rigid structural assumptions of traditional choice models that define expenditure elasticity and the restrictive utility functional forms these models often impose. By introducing Gaussian process priors on utility functions, we provide a flexible, utility - based model for understanding expenditure - driven changes in consumer choices. We demonstrate that relaxing the functional form on the outside good within the framework of constrained utility maximization leads to more flexible substitution patterns. This has implications for understanding preference for variety and quality. This methodological advance enables the model to capture non-linear rates of satiation and precise baseline preferences—details that traditional non - homothetic (i.e., expenditure-variant preferences) parametric models often overlook due to their assumptions of a given utility functional form. Through its automatic detection of non-linear consumption patterns from the data, the model provides more flexible statistical inference, offering valuable theoretical and practical insights for improved pricing decisions. 

Chapter 3, titled “Digital Twins: A Generative Approach for Counterfactual Customer Analytics”, proposes an innovative methodology to optimize customer surveys in a competitive landscape. Leveraging a unique dataset of quarterly cross-sectional survey responses from major U.S. telecommunications providers from 2020 to 2022, this paper introduces the concept of ‘Digital Marketing Twins.’ These are generative vi models of customer preferences that provide counterfactual responses under differ?ent scenarios. Here, the concept of “generative model” means that I explicitly give the sequence of steps describing how the data were created, i.e., the data generating process, including unknown model parameters. The methodology uses a novel deep generative and probabilistic latent factor model, which captures individual - level brand affinity for each brand and time period, accounting for observed het?erogeneity and firm-side factors. Utilizing Bayesian optimization, the model offers individual-level marketing action recommendations. It shows promising results in identifying marketing actions most likely to increase customer satisfaction, offering a “path of least resistance” at the individual level.