Skip to content

Recommendation systems

Recommendation systems are algorithms that suggest content, users, communities, and other items to platform users based on their interaction history, stated preferences, or similarity to other users. On social media platforms, these systems drive what content appears in feeds, notifications, and search results, and thus play a direct role in determining which information reaches users and how it spreads.

How recommendation systems work

Modern social media recommendation pipelines typically include:

  • Candidate sourcing — A large pool of candidate items (tweets, videos, posts) from multiple sources (user's follows, out-of-network sources identified through embeddings or graph-based retrieval).

  • Light ranker — A fast, simple model (e.g., logistic regression) that filters the candidate pool to a manageable size using basic features.

  • Heavy ranker — A more complex neural network (e.g., MaskNet, transformers) trained to predict the probability of user engagement for multiple action types (likes, shares, comments, watch time).

  • Score aggregation — Combining predicted engagement probabilities into a final ranking score; typically a weighted sum of predictions across engagement types.

The specific design of each component—especially the aggregation layer—determines what content gets amplified and what gets buried.

Relation to misinformation propagation

Recommendation systems play a central role in determining whether misinformation propagates faster than accurate information:

  • Engagement optimization — When systems optimize for engagement metrics (likes, shares, comments, watch time), they often inadvertently amplify false claims, which tend to provoke emotional, reflexive responses.

  • Engagement composition — Different engagement types may reflect different cognitive processes. Why Does Misinformation Propagate Faster? An Algorithmic Perspective on X shows that X's score-aggregation layer treats instant reactions (likes, retweets) and deliberative responses (replies, quotes) as fungible evidence of quality, allowing low-credibility content that draws disproportionately from fast engagement to receive amplification advantages.

Key papers