Social Media Feed Elicitation

Paper Title

Social Media Feed Elicitation

Publication Info

• Topic area: User-centered customization of social media feeds using AI-driven elicitation.
• Keywords: social media, feed algorithms, user preferences, large language models, feed customization, elicitation interviews, personalization, user agency, content moderation, feed ranking.

Background and Problem

• Problem / challenge: Social media feed algorithms are opaque and engagement-optimized, leading to user dissatisfaction and various societal harms. While tools for feed customization exist, users struggle to articulate their preferences comprehensively, resulting in suboptimal feeds.
• Significance: Addressing this issue empowers users to control their social media experiences, potentially mitigating issues like polarization, misinformation, and emotional distress.
• Motivation and related work: Prior work has focused on moderation tools, chronological feeds, and value-driven ranking. However, users face an "articulation gap" where they cannot fully specify their ideal feed due to cognitive biases like the planning fallacy and the illusion of explanatory depth. This paper addresses this gap by proposing a structured elicitation process.

Solution

• Proposed approach: Feed elicitation interviews—an interactive, LLM-powered method that guides users through specifying their preferences for custom social media feeds.
• Novelty:
  1. Identification of the "articulation gap" as a key barrier to user-driven feed customization.
  2. Development of an LLM-powered interview process to scaffold user preferences.
  3. Integration of elicitation with a feed-generation system that ranks posts based on user-defined criteria.
  4. Empirical evidence demonstrating the effectiveness of the approach in improving feed quality and user satisfaction.
• Procedure and key techniques:
  1. Elicitation interview: Conducted in three stages—(1) defining the purpose of the feed, (2) identifying relevant topics, and (3) specifying post attributes (e.g., tone, framing).
  2. Preference synthesis: User responses are synthesized into a detailed feed specification.
  3. Feed creation: Posts are filtered and ranked using a two-step LLM-based classification system—relevance filtering followed by quality ranking.
  4. Evaluation: A between-subjects study (N = 400) comparing baseline manual feed specifications with those produced by the elicitation interview.

Results

• Concrete findings:
  • 48% of participants preferred feeds generated via the elicitation interview, compared to 29% preferring the baseline and 23% neutral.
  • Elicited feeds had a higher proportion of approved posts (56% vs. 50% for baseline) and fewer disapproved posts (22% vs. 28% for baseline).
  • Structured manual descriptions (non-interactive) showed no significant improvement over the baseline.
• Advantage over baselines:
  • Elicitation interviews led to more personalized and actionable feeds, addressing user preferences more comprehensively.
  • Participants appreciated the removal of irrelevant or disruptive content and the inclusion of nuanced preferences.
• Experiments / evaluation:
  • Participants rated 40 posts (20 per feed) on a 7-point scale and expressed overall feed preference.
  • NASA-TLX metrics assessed cognitive effort, frustration, and satisfaction.
  • Statistical analyses included Wilcoxon signed-rank tests and cumulative logit mixed-effects models.
• Limitations and future work:
  • Overspecification in elicitation sometimes led to overly narrow feeds with low-quality content due to limited post inventory.
  • The system struggled to capture the relative prioritization of user preferences.
  • Future work should explore larger post inventories, dynamic prioritization, and longitudinal impacts on user well-being.

Summary

This paper introduces feed elicitation interviews, an LLM-powered method to guide users in articulating their preferences for custom social media feeds. The approach addresses the "articulation gap" by prompting users to consider purpose, topics, and post attributes in a structured manner. Evaluation shows that elicited feeds are preferred over baseline manual specifications, with higher approval rates and better alignment with user desires. While the system effectively enhances user agency, challenges like overspecification and prioritization remain. This work advances tools for user-driven feed customization and has broader implications for elicitation in other domains.