Unpacking the Eye of the Beholder: Social Location, Identity, and the Moving Target of Political Perspectives

Technical Summary: Unpacking the Eye of the Beholder

Problem Statement
Computational social science has increasingly adopted machine learning tools to analyze political text and imagery at scale. However, these tools are predominantly built on a paradigm of "ground truth absolutism," which assumes that a single correct label exists for any data instance, independent of the observer. In evaluative tasks—such as determining the sentiment or perceived violence of a political image—this assumption is flawed. Political science theory establishes that evaluation is mediated by social location and identity; where a person stands politically and socially determines what they see. Standard computational approaches treat the disagreement between annotators of different identities as "noise" to be eliminated via majority voting. By averaging these disagreements away, existing tools discard the very phenomenon political science cares about: the structured lines of tension and evaluative conflict that define political disagreement. Consequently, current models produce single scores that treat a piece of media as having a fixed meaning for everyone, failing to capture how political images function as "moving targets" dependent on the audience.

Methodology: The Perspectivist Visual Political Sentiment (PVPS) Classifier
To address this, the author develops the Perspectivist Visual Political Sentiment (PVPS) classifier. Unlike standard classifiers that predict a single sentiment score, PVPS predicts an evaluative profile for each image, recording which specific audience segments rate the image more favorably and along which identity lines divergence occurs.

• Data: The model is trained on approximately 82,000 evaluations of 1,264 political images provided by 5,575 U.S. adults across four survey waves (2022–2025). Respondents provided ratings on a 7-point attitude scale alongside full demographic batteries (party, ideology, age, gender, education, income, ethnicity). A supplementary dataset of 7,543 images annotated by over 2,000 workers was incorporated via semi-supervised self-training.
• Architecture: The pipeline extracts 1,666 feature vectors per image from six sources:
  1. CLIP (512 dims) and DINOv2 (768 dims) for visual content and spatial layout.
  2. Gemini captions converted to TF-IDF vectors (300 dims).
  3. Semantic features (50 dims) indicating the presence of predictive words.
  4. Concept features (24 dims) for political objects (e.g., police, flags, signs).
  5. Political topic features (12 dims) for policy domains.
     These vectors are compressed through a projection layer and two residual blocks into a 128-dimensional representation, which is then mapped to class probabilities (Consensus vs. Divergence) via a softmax layer.
• Training Strategy: The system trains separate classifiers for 112 distinct axes, covering primary political divides (Party, Ideology, Thermometer), sociodemographic axes (Age, Gender, Education, Income, Hispanic), and their intersections (e.g., Young Democrat vs. Old Republican).
  • For political axes, the model learns to predict group divergence directly from visual content.
  • For sociodemographic axes, where per-image gaps are noisy, an extended pipeline (Path B) is used. This path predicts individual ratings based on image features and respondent attributes, then computes the gap between prototypical respondents, combining this with the visual signal (Path A) via a meta-learner.
• Validation: An ensemble of ten classifiers trained on different data splits and random seeds is used. Performance is measured by the median minimum per-class accuracy to ensure robustness against class imbalance.

Key Results

• Classification Performance: The classifier successfully predicts evaluative divergence on all three primary political axes (Party, Ideology, Thermometer) with median minimum-class accuracies ranging from 68.7% to 78.9%. It also achieves high accuracy (60–83%) on many cross-party and intersectional axes.
• Sociodemographic Signals: Pure sociodemographic axes (e.g., Age, Gender) show weaker signals when relying solely on image-level disagreement (50–58% accuracy). However, the extended pipeline, which conditions on respondent attributes, recovers significant signal for Age (68.9%), Gender (62.0%), and Education (61.0%), demonstrating that social identity shapes evaluation but often in ways entangled with political structure.
• Reanalysis of Existing Studies:
  • Casas and Webb Williams (2019): Applying PVPS to Black Lives Matter (BLM) imagery reveals that the mobilizing effect of emotions is not uniform. Enthusiasm mobilizes sharing for images favorable to Democratic women, while fear and disgust drive engagement for images favorable to Republican men. Averaged emotion scores obscure these divergent mechanisms.
  • Won, Steinert-Threlkeld, and Joo (2017): The study finds that images rated as "violent" by annotators are the same images predicted by PVPS to be favorably evaluated by Republicans, conservatives, and Republican-leaning respondents. Conversely, images with protest signs (associated with lower perceived violence in the original study) are predicted to be favored by Democrats. This suggests perceived violence is not a neutral visual attribute but follows partisan fault lines.

Significance and Claims
The paper claims two primary contributions that bridge political science and computational methods:

1. For Political Science: It provides a scalable method to measure group-mediated evaluation. While theory has long established that political judgment is identity-sorted and intersectional, empirical tools have lagged, often collapsing disagreement into majority labels. PVPS makes structured disagreement observable and predictable, allowing researchers to measure how visual persuasion effects vary across specific audience segments rather than assuming a monolithic audience response.
2. For Computational Political Science: It demonstrates that the axes of group conflict over political images carry a signal strong enough to be modeled directly from visual content. By incorporating perspectivist assumptions (that disagreement is signal, not noise) into the model architecture, the interpretation of how images are received changes substantively. This allows for the reanalysis of media effects and visual persuasion at scale, acknowledging that "what a political image conveys" is a moving target dependent on the viewer's social location.

The paper concludes that measuring political images requires knowing "whom it is moving." It does not claim to solve all issues of bias or generalizability (noting the U.S. context is an outlier in affective polarization), but it establishes a framework where the "eye of the beholder" is no longer a source of error to be averaged out, but the central object of analysis.