FactorFlow: A Visual Analytics Workspace with Large Language Model-Assisted Interpretation for Factor Analysis

Abstract

In exploratory factor analysis (EFA), one typically aims to extract and describe a small number of factors (i.e., latent variables) based on the relationships among numerous manifest variables (i.e., directly observable variables). In practice, performing EFA entails examining different factor models (and rotations) to identify the underlying latent structure. Now, the primary criterion for evaluating a factor model is interpretability. That is, the preferred model is the one that yields a meaningful, coherent, and theoretically defensible factor structure. However, gauging a model's interpretability is not a trivial task, as it is subjective and often requires keeping track of large amounts of information simultaneously. Because of this, researchers typically employ various visualizations to interpret models and determine the "best" one. Hence, we introduce FactorFlow, a visual analytics workspace for performing EFA end-to-end. Using FactorFlow, one can fit and rotate factor models, perform model diagnostics, and more. The main component of the tool is a dashboard with a comprehensive set of interactive visualizations, where a user can easily dissect a factor model and even compare two models side-by-side at the same time. Moreover, several large language models are integrated with FactorFlow, enabling the user to generate and assess automated factor interpretations written in natural language. With multiple views and readily available calculations, FactorFlow can enable the researcher to efficiently and effectively understand factors and ultimately, perform EFA. Finally, we conducted a usability study to identify strengths and weaknesses, and capture feedback to incorporate in the app.