Click-through rate

Instagram’s click predictions can concern a specific action: Mosseri described predicting whether someone will tap a Feed post author’s profile picture3. He listed commenting, liking and sharing as separate outcomes Instagram predicts3. Mosseri said Instagram makes this series of predictions for each Feed post3.

Meta’s Kunlun research describes estimating click likelihood using a person’s sequence of past interactions with items1. Meta’s research also considers information about the person and the item, including demographic information and ad details1. In Meta’s research, these inputs serve different purposes: relatively stable contextual information represents explicit preferences, while interaction history helps represent interests that change over time1.

Meta’s LLaTTE research describes using predicted click probabilities both to order ads and to inform auction pricing4. Meta’s research considers click predictions alongside predictions of conversions and reports results for Facebook Feed and Reels advertising, rather than Instagram’s ordinary posts4.

Meta’s engineering team describes balancing Instagram notification click performance with variety by lowering the priority of notifications too similar to recently delivered ones, with adjustable controls for content, author and product type5. The team reported that this approach reduced daily notification volume while improving CTR5. A Meta patent concerning recommended search cards proposes combining predicted CTR with the value of the resulting activity, including user interest, satisfaction or monetary value, allowing a less frequently clicked activity to receive a higher value6.

Meta’s engineering team describes checking Instagram’s click predictions through calibration, which asks whether a system predicts more or fewer actions than actually occur2. The team calculates calibration as predicted CTR divided by observed CTR, with a perfect predictor centered on one2. Meta’s engineering team also uses a measure called normalized entropy to assess a different question: how well the predictor distinguishes cases where someone acts from cases where they do not2.