Multi-stage ranking

Instagram Search illustrates multi-stage ranking by first collecting possible results relevant to the search words, including hashtags, places, reels, posts, profiles and audio1. It scores those results using factors such as content type, how well they match what the person is searching for, and what that person tends to interact with1. Search then narrows the selection, applies checks for content integrity and duplicate results, and orders the remaining results by score1. Explore gathers up to 1,500 photos and videos, then compares characteristics of the content with characteristics of the person viewing it to select the most promising possibilities2. The Instagram Reels Chaining system gathers possible reels, including those from followed accounts, and uses a computer system requiring less computing work to select approximately 100 based on information such as reel length, similarity to other reels and the viewer’s interactions3. Further computer systems then estimate which of those reels the person will find most relevant and valuable3.

Meta’s patent on multi-stage ranking describes a proposed way to allocate computing effort, rather than establishing how Instagram currently operates6. In Meta’s patent, the selection steps use computer systems that estimate how likely someone is to interact with each content item6. Meta’s patent evaluates possible adjustments to those steps and their predictions by calculating both the computing time required and a measure of the expected value of the selected content to the system6. Within a given computing-time budget, the patent’s method chooses adjustments intended to increase that expected value6.