Problem
Instagram ranks notification candidates with ML models trained for positive engagement such as click-through rate (CTR) and time spent. Such models overprioritize the authors and product types a person has engaged with before. The result is overexposure: notifications about one friend only, or only Stories while the person would also like Reels or Feed. Over time notifications feel spammy and people may disable them.
Example from the article: a person who often interacts with one friend's content gets notifications about that friend only. A person who recently engaged with Stories more than Reels gets only Story notifications.
Solution in brief
Add a diversity layer on top of the existing engagement models. For each candidate it measures similarity to recently sent notifications along several dimensions and multiplies the base score by a penalty between 0 and 1. The adjusted scores re-rank candidates, and the quality bar picks the top one.
Questions this case answers
- How to add controllable diversity on top of an engagement-optimized ranker that overexposes what a user already interacted with, without retraining the ranker
- How to stop an engagement-optimized notification model from sending the same author or the same product surface again and again?
- How to add diversity to notification ranking without losing personalization and relevance?
- How to tune the strength of repetition penalties separately for author, content, notification type and product surface?
Tools and methods
- Diversity layer — post-processing on top of the engagement models; multiplies each candidate's score by a demotion factor
- Maximal marginal relevance (MMR) — similarity signal of a candidate to the history of sent notifications
- Per-dimension demotion weights — adjustable strength of the penalty for author, content, notification type and product surface
- Quality bar — final gate that sends the top-ranked candidate passing both the ranking and diversity criteria
Google or a one-time email code. No password.