Embedding-based retrieval

Instagram Explore gathers up to 1,500 photos and videos for consideration.1 Meta’s Transparency Center lists two-tower sourcing as one of several methods used to assemble this collection.1 The collection includes media from authors a person has engaged with and media similar to content that person has engaged with.1 Explore’s early scoring model calculates similarity using information about both the person and the media, producing a shortlist of 100 items.1

Meta’s patent describes one approach to building a person’s numerical representation from things they have previously interacted with.3 In Meta’s patented approach, those previously encountered things have numerical representations of their own, which are combined to construct the person’s representation.3 Meta’s patent specifies a weighted combination, meaning that some interactions count more than others when constructing that representation.3

Meta’s patent uses similarity measurements between a person’s representation and the representations of available items to put those items in order and select recommendations, which can include content or pages.3 Meta’s SilverTorch research explains the scale problem behind this kind of selection: scoring every available item at the time of a request is impractical, so an initial search narrows the collection to thousands of candidates.2 Meta’s research describes approximate nearest-neighbor search, a way to find likely close matches among numerical representations without scoring the entire collection individually.2 Meta’s research also describes filtering that removes candidates that do not meet requirements such as language or eligibility, meaning whether an item is allowed to be shown to that person.2