Topics · How it works
Cold-start recommendations
For Instagram creators, cold-start recommendations concern suggesting content with little interaction history, but Meta’s research and patent describe approaches outside documented Instagram behavior123. Meta’s Facebook Marketplace research addresses short-lived listings with little activity, while a Meta patent proposes recommending content that has received relatively little interaction13. Meta’s recommendation research also examines trying different plausible predictions of audience response rather than relying on one fixed estimate2. These approaches do not establish Instagram-specific controls or effects on creators’ reach123. The latest work discussed is Meta’s November 2024 research on Epinet for cold-start recommendations, not guidance for Instagram creators2.
Finding matches with limited feedback
Meta research and patents. These may not describe how Instagram works today.
Meta’s Facebook Marketplace research examines cold-start recommendations for product listings whose short lives make it difficult to gather useful activity on each listing1. Meta’s research describes using shopper information alongside product photos and descriptions to find a shortlist of nearby products, which a separate system then orders for presentation1. A Meta patent describes another matching approach: examining the profiles of people who initially watched or interacted with newly obtained content, then finding other people with similar profiles who might be interested3.
Testing audience response to new content
Meta research and patents. These may not describe how Instagram works today.
Meta’s November 2024 Facebook Reels research examines cold-start recommendations through experiments with a method called Epinet for new content that has been shown few times2. Meta’s research describes choosing recommendations by trying different plausible predictions of audience response rather than relying on one fixed estimate2. Meta’s research balances trying content whose response is less certain against selecting content expected to perform well2. In Facebook Reels experiments, Meta’s research reported improvements involving likes and completed video views2.