Value Alignment of Social Media Ranking Algorithms

Abstract

While social media feed rankings are primarily driven by engagement signals rather than any explicit value system, the resulting algorithmic feeds are not value-neutral: engagement may prioritize specific individualistic values. This paper presents an approach for social media feed value alignment. We adopt Schwartz's theory of Basic Human Values -- a broad set of human values that articulates complementary and opposing values forming the building blocks of many cultures -- and we implement an algorithmic approach that models and then ranks feeds by expressions of Schwartz's values in social media posts. Our approach enables controls where users can express weights on their desired values, combining these weights and post value expressions into a ranking that respects users' articulated trade-offs. Through controlled experiments (N=141 and N=250), we demonstrate that users can use these controls to architect feeds reflecting their desired values. Across users, value-ranked feeds align with personal values, diverging substantially from existing engagement-driven feeds.

Figures (9)

• Figure 1: Visualization of Schwartz's 19 Basic Human Values. Values close to each other in the circle tend to share similar motivations whereas opposing values often are in tension with each other. The 19 values are also organized into four broader groups: self-transcendence, conservation, self-enhancement, and openness to change.
• Figure 2: Our method integrates multiple values together when ranking content. We show the top posts from the same inventory of tweets but ranked by two different sets of value weights. The left shows a feed that prioritizes "Achievement" and "Dominance" (Power) while downranking "Humility. The right feed prioritizes "Humility" while downranking "Face" (Reputation).
• Figure 3: The experimental platform renders feeds emulating the style of Twitter. When comparing two feeds for recognizability, the engagement and value-ranked feed are shown side-by-side on the screen with generic labels (e.g., "Feed A") at the top of each feed.
• Figure 4: On our experimental platform, participants are able to re-rank their feed by adjusting value sliders. Participants see a single feed on the right-side of their screen presented with the sliders that they can use to re-rank the content on the left-side. While participants are shown all available sliders, the number that they can adjust (i.e., 1, 2, 3, 4, 5, or all) depends on their assigned condition.
• Figure 5: Recognizability drops as the number of sliders changed increases but remains above random chance even when participants are able to change all 19 sliders. We group together conditions with more than 5 sliders changed into a single "5+" category due to the small number of participants in higher-change conditions. All error bars represent 95% confidence intervals.