Recommendation scoring

Instagram's documented Feed and Reels processes narrow the available content before calculating final relevance scores23. For Feed posts from followed accounts, a lightweight model—a simpler automated prediction system that narrows the posts considered—selects approximately 700 posts using information such as who created a post and its content type2. The later Feed relevance-scoring stage covers about 500 posts2. The Reels process uses a lightweight model to select approximately 100 reels, considering their length, similarity to other reels and fit with content the viewer tends to interact with3. Before final Reels scoring, integrity checks remove content that cannot be recommended or may go against Community Guidelines3.

For suggested Feed posts, Instagram's relevance predictions draw on a viewer's interests, engagement with similar content and previously reshared posts1. Meta's Transparency Center also lists an assessment of a post's content quality generated entirely by an AI language model1. Meta's August 2023 patent proposes scoring content by matching its characteristics with a person's attributes and interests, but does not establish that Instagram uses this method4.

Meta's October 2019 patent proposes a social-network recommendation method, rather than documenting an Instagram feature5. Meta's patented approach finds other items saved by people who saved the same item and considers how unusually often those other items are saved5. To calculate relevance, Meta's patented method supplies both a possible content item and the viewer's context to an automated scoring system, which uses content people save and later view5. Meta's patent then uses those scores to select other content to present to the viewer5.

Meta's patents describe ways to combine several recommendation scores, without establishing that Instagram uses these methods67. Meta's November 2022 patent describes a virtual assistant that merges two sets of scores into a total for each possible recommendation, then selects suggestions from the resulting ordered list6. When a recommendation lacks a score in one set, Meta's patent allows alternatives such as assigning zero or copying its other score when the two sets use a comparable scale6. Meta's November 2021 patent describes a different approach: one automated prediction system scores a first set of content characteristics, while another scores groups of similar characteristics specific to the content provider7. A third prediction system in Meta's patented approach combines those two scores into a final score used to select content for presentation7.