Date of Award
Spring 2023
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Management
First Advisor
Sudhir, K
Abstract
Marketing models typically focus on how structured product characteristics impact consumer preferences. However, visual characteristics of products present in unstructured image data play an important role in impacting preferences for many categories. In this dissertation, we develop a method to automatically discover and quantify visual characteristics from product images using a disentanglement-based approach. Disentanglement aims to identify statistically independent and human interpretable characteristics, so a change in one visual characteristic produces a change in the visual space only along one human interpretable visual characteristic. The existing deep learning literature has shown that it is impossible to obtain unique representations of disentangled visual characteristics without supervision. However, since the goal is to discover human interpretable visual characteristics, this presents a challenge. In the first chapter, we focus on using readily available structured product characteristics as supervisory signals. We apply this method to automatically discover visual product characteristics of watches, and discover 6 semantically interpretable visual characteristics providing a disentangled representation. We find that while some supervisory signals such as 'brand' help in finding visual characteristics as compared with an unsupervised method, other signals such as 'price' do not. In the second chapter, we leverage consumer preferences on human interpretable visual characteristics obtained from market outcome data to use as supervisory signals. First, we obtain product fixed effects from a supply and demand model of market equilibrium. These product fixed effects subsume consumer preferences on visual characteristics apart from other unobserved characteristics. Next, we train the disentanglement algorithm to find visual characteristics that are not only independent and human interpretable but can also predict the product fixed effects. We find that including visual characteristics significantly improves prediction of market outcomes. In the third chapter, we apply these visual characteristics in two marketing applications. First, we show visual characteristics can be used to implement choice-based visual conjoint analysis. Second, we show how human interpretable visual characteristics can be used to make market structure maps. We show how two products similar in the structured product characteristic space can be far apart in the visual characteristic space, and discuss the implications for competitive product positioning.
Recommended Citation
Sisodia, Ankit, "Quantifying Visual Characteristics of Products" (2023). Yale Graduate School of Arts and Sciences Dissertations. 1061.
https://elischolar.library.yale.edu/gsas_dissertations/1061
