Engagement prediction

For comparison with Instagram’s predictions about Reels, Meta’s ads research describes a prediction model—a computer system trained to estimate likely responses—that produces separate probabilities for clicking an ad or completing a desired action12. A Meta patent describes another target for prediction: whether a product or combination of products shown in posts will prompt a positive reaction or otherwise engage viewers3. These examples concern Meta’s ads research and a patented product system, rather than establishing which individual responses Instagram currently predicts for Reels23.

Meta’s ads research shows how engagement predictions can draw on information about the person, the ad, the person–ad pairing, and the circumstances of the request to show an ad2. For a documented example outside Instagram Reels, Meta’s Transparency Center says Threads considers previous engagement with similar accounts and content, including a viewer’s engagement over the past week and time spent viewing posts4. The Meta product-engagement patent bases its product scores on tracked interactions, connecting its estimates to observed viewer behavior3.

Engagement predictions can inform content ordering, although the documented Threads system and Meta’s patents and Facebook research describe different approaches rather than one universal Instagram formula456. Meta’s Transparency Center says Threads places content higher in the feed when its system predicts that it will provide more value to the viewer4. A Meta feed patent describes adding values associated with predicted actions, such as liking, hiding, sharing, commenting or viewing, while reducing some actions’ influence; the resulting score can affect both whether an item appears and its position6. For Facebook Reels, Meta’s engineering team describes predicting video engagement after someone sees a friend bubble—a highlight showing a friend’s reaction—and using adjustable weights, meaning different amounts of influence, to balance video engagement against social interaction5.

Meta’s research on a memory layer extends the prediction-to-ranking process by examining whether predictions remain available and consistent when a recommendation system is used, rather than documenting an Instagram-wide deployment7. Meta’s research describes problems when an item is missing from a cache, a store of previously prepared information about content, and reports a 5–12% gap between prediction-performance measurements during training and live use across engagement-prediction tasks7. Meta’s research proposes training and running the system with the same trainable cache so that both stages use fresher, more consistent content information7. Meta’s research also reports that this approach provides a prediction for every item, including newly created Reels, instead of leaving items out because their information is missing from the cache7.