Value Alignment of Social Media Ranking Algorithms

Paper Title

Value Alignment of Social Media Ranking Algorithms

Publication Info

• Topic area: Value-based ranking in social media feed algorithms.
• Keywords: Value alignment, social media, feed ranking, Schwartz Basic Values, large language models, user control, engagement algorithms, value trade-offs, algorithmic personalization, human-computer interaction.

Background and Problem

• Problem / challenge: Social media feed algorithms prioritize engagement metrics (e.g., clicks, likes) without considering the values embedded in the content they amplify. This approach disproportionately emphasizes short-term individualistic values, potentially leading to societal harms like polarization and misinformation.
• Significance: Algorithmic feeds influence attitudes, beliefs, and behaviors, making it critical to explore how values can be intentionally incorporated into feed rankings to align with users’ personal and societal goals.
• Motivation and related work: Previous work has focused on mitigating misinformation or political content but lacks a comprehensive framework for incorporating a broad set of human values into ranking algorithms. Schwartz’s theory of Basic Human Values offers a structured and empirically validated system for addressing this gap.

Solution

• Proposed approach: A value-aligned ranking algorithm that uses Schwartz’s Basic Human Values to classify and rank social media content based on user-specified value preferences.
• Novelty:
  1. A framework for integrating a comprehensive value system into feed ranking algorithms, operationalized using Schwartz’s Basic Human Values.
  2. Two controlled experiments demonstrating that users can recognize and configure value-aligned feeds.
  3. Empirical insights into how users select values and how these selections reshape their feeds.
• Procedure and key techniques:
  1. Classify social media posts using Schwartz’s 19 values via a large language model (LLM).
  2. Allow users to specify value weights (ranging from -1 to 1) through a slider interface or a questionnaire.
  3. Re-rank feeds by computing a weighted score for each post based on the user’s value preferences.
  4. Validate the approach through experiments comparing value-aligned feeds to engagement-based feeds.

Results

• Concrete findings:
  • In Study 1 (N = 141), participants identified value-aligned feeds in 76.1% of trials, significantly above random chance.
  • In Study 2 (N = 250), recognizability declined as more values were activated but remained above random chance (63.4% overall).
  • Engagement feeds prioritize individualistic values like “Stimulation” and “Hedonism,” while value-aligned feeds amplify societal values like “Caring” and “Universal Concern.”
• Advantage over baselines:
  • Value-aligned feeds diverged significantly from engagement-based feeds, with a mean Kendall’s τ of 0.06, indicating minimal overlap in ranking.
  • LLM-generated value labels aligned more closely with consensus annotations than individual human annotators (LLM-Consensus MAE = 0.95 vs. Human-Consensus MAE = 1.07).
• Experiments / evaluation:
  • Study 1: Participants compared engagement feeds to single-value-aligned feeds using their own Twitter content.
  • Study 2: Participants adjusted up to 19 value sliders to create multi-value-aligned feeds and compared them to engagement feeds.
  • Metrics: Recognizability of value alignment, task load, and changes in value strength in re-ranked feeds.
• Limitations and future work:
  • Limited to U.S.-based participants; results may not generalize across cultures.
  • Short-term experiments; longitudinal studies are needed to assess long-term effects of value-aligned feeds.
  • Current method does not account for contextual factors like the author of a post.
  • Future work could explore non-linear ranking functions, natural language interfaces for value elicitation, and culturally specific value systems.

Summary

This paper introduces a method for aligning social media feed rankings with user-specified values based on Schwartz’s Basic Human Values. Using large language models, the approach classifies and ranks content by value expressions, enabling users to reconfigure their feeds through a slider interface. Two experiments validate that users can recognize and configure value-aligned feeds, which diverge significantly from engagement-based rankings. The findings highlight the potential for value-based personalization to empower users and mitigate the societal harms of engagement-driven algorithms, while identifying challenges in scaling and generalizing the approach.