A Machine Learning Model for Understanding How Users Value Designs

Thesis Advisor: Terry Knight
Reader: Stefanie Mueller

People value furniture designs in subjective and multifaceted ways. Understanding these nuances helps designers develop intuition for creating new designs. And yet, while an abundance of furniture designs already exist, our ability to access and leverage data about how users value them is limited. If our goal is to continually improve the outcome of how users value furniture designs—whether through design practice or commerce—we need a framework for collecting and interpreting user feedback at scale.

In this thesis, I demonstrate a number of advances toward developing a machine learning (ML) model of how designs are valued. The model can be used to better understand the implications of furniture design decisions, as well as for commercial strategy.

Existing ML systems have been trained on the physical and aesthetic features of completed designs. These top-down methods do not capture the nuances of how users actually value the various functions of their furniture. To improve on these methods, I first develop a framework for ingesting and classifying user feedback about how designs are valued.  Next, I conduct a user survey to test this framework, generating a bottom-up, labeled dataset which requires no post-processing. Finally, I develop a framework for the computational analysis of this data. The framework is based on a probabilistic ML model trained on the real user data collected.

Through visualizations, I demonstrate how the ML model can be used to show an overview of large data sets of user feedback, model the preferences of new users and augment existing data sets. The model’s efficacy is evaluated by comparing results to a test data set split from the original data set.

This framework represents a step toward a future in which datasets for furniture—and other design domains—are more accessible. By making user feedback available to designers at scale, and establishing methods for collecting this data, we can accelerate the development of designer intuition and deliver significantly greater value to users.