For more about my research and for my most up-to-date CV, please visit my personal website: www.rtdew.com

My current research focuses on how managers can harness cutting edge statistical and machine learning methodologies to improve decision making, particularly in the area of marketing analytics. I am also interested in the analysis of images, and incorporating visual data in traditional marketing models.

Publications and working papers:

• Bayesian Nonparametric Customer Base Analysis with Model-based Visualizations
  Ryan Dew and Asim Ansari
  Forthcoming at Marketing Science. Working paper available on SSRN.
  
  
• Dynamic Preference Heterogeneity
  Ryan Dew, Yang Li, and Asim Ansari
  Revision invited at Journal of Marketing Research. Working paper available on SSRN.
  
  

Works in progress:

• Letting Logos Speak: Deep, Probabilistic Models for Logo Design
  In this work, we investigate how brand identity is manifest in the features of logos, by decomposing logos into data through image processing and computer vision techniques. The goals of the project are both descriptive and prescriptive: what brand constructs are associated with which features of logos? Can a computer generate a typical logo, given a brand description? By representing logos as data, we can address these questions and more in a data-driven fashion, using sophisticated statistical techniques.
  
  
• Customer-Centric Data Fusion
  Marketers face a deluge of data on all aspects of their customers, including what they are buying, how they are using their products, and what they are saying about their purchases. We seek to address the problem of understanding general patterns across many modes of customer behavior, by modeling consumers through a set of latent, inferred traits that are related to behaviors across all domains of interest. Our customer-centric approach allows firms to leverage data across domains, even in the presence of missing data, to better understand and predict customer behavior in any particular domain, using existing probabilistic models of customer behavior.
  
  
• Scalable Decision Support Systems for Robust CRM
  Modern machine learning methods provide a powerful basis on which to build decision support systems for customer relationship management, yet many of the most flexible and comprehensive of these methods are computationally complex or slow to estimate on very large datasets. In particular, hierarchical models that capture the individual-level heterogeneity that is crucial for CRM and targeted marketing can suffer from scalability problems. In this work, we develop and apply methods for scalable, approximate inference using stochastic gradient methods, that allows robust and hierarchical decision support systems, such as the recently developed Gaussian process propensity model (GPPM), to be scaled to very large datasets. In particular, we show that use of stochastic gradient inference methods allows accurate and fast estimation of complex CRM models like the GPPM, which in turn facilitates two previously difficult forms of analysis: first, models can be fit to more subsets of the customer base, allowing for nuanced understanding of differences across acquisition channel and customer characteristics. Second, the model can be automatically adjusted to new settings, where more expressive structure is needed to understand customer spending behavior.