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. We seek to automatically discover and quantify visual characteristics (attributes) from image data using a disentanglement-based approach. While the deep learning literature has shown that supervision is required to obtain unique disentangled representations, ground truth visual characteristics are typically unknown. We develop a method that does not require such supervision, and instead uses readily available structured product characteristics as supervisory signals to enable disentanglement. Our method does not need prior knowledge of characteristics, yet we are able to discover semantically interpretable and statistically independent characteristics. Moreover, the method quantifies the levels of each discovered product characteristic, necessary for managerial tasks such as demand modeling and conjoint analysis. We apply this method to automatically discover visual product characteristics of watches, and discover 6 semantically interpretable visual characteristics providing a disentangled representation. Our results find the supervisory signal `brand' best promotes disentanglement relative to an unsupervised approach. We lastly demonstrate how consumers preferences may be assessed over these discovered visual characteristics using a choice-based conjoint analysis.

This publication has 44 references indexed in Scilit: