Computer Vision with Blindfolds: Estimating Ideology on Social Media from Observed User Behavior – Matias Piqueras

Ideology, understood as a low-dimensional representation of political preferences, is one of the most important constructs in political science. Despite its ubiquity, ideology remains difficult to measure empirically at scale. Yet, being able to do so, is increasingly relevant in light of the proliferation of online platforms, where researchers are often interest in how the political debate is shaped, contested and consumed along ideological lines. We introduce the Poisson Factorization Ideal Point (PFIP) model, a scalable Bayesian framework for estimating latent ideological positions from large-scale behavioral data. Building on hierarchical Poisson factorization, PFIP incorporates a kernel approximation to ideological distance, enabling the recovery of interpretable ideal points on a symmetric, continuous scale while maintaining computational efficiency. This approach addresses two persistent challenges in ideal point estimation: scalability to billions of observations and robustness to confounding factors. We validate the model by replicating canonical results from roll-call votes, survey responses and social media data, demonstrating that PFIP produces estimates comparable to established methods while accommodating diverse data types (counts, binary, multinomial). We then apply the model to millions of comments on visual political content, generating a large-scale ideological mapping of accounts and content. Validation against external ideological benchmarks for media outlets, channels, Podcasts, and members of Congress shows strong convergence.

About Matias Piqueras