Why Notification Diversity Matters

Notifications are a primary way Instagram draws people back to the app, but the models that decide what to send are usually trained to maximize engagement signals like click-through rate (CTR). That optimization can create a feedback loop: the system keeps surfacing the same authors and product surfaces that already performed well, even when other content could be just as relevant.

This overexposure typically shows up in two patterns. First, users may get repeated notifications about the same friend or creator, even if they engage with others too. Second, notifications may cluster around one product surface, such as Stories, while ignoring Feed or Reels content the user also cares about. Over time, this repetition makes notifications feel spammy and can push users to disable them entirely.

A Diversity Layer on Top of Engagement Ranking

Instagram’s new ranking framework adds a diversity layer on top of existing engagement ML models. The layer evaluates each notification candidate’s similarity to recently sent notifications across dimensions like content, author, notification type, and product surface. Candidates judged too similar receive a multiplicative penalty—expressed as a demotion factor—before re-ranking. The final selection is the top candidate that clears both the relevance and diversity thresholds.

Figure.1: Instagram’s diversity-aware ranking framework where the diversity layer sits on top of the existing modeling layer and penalizes notifications that are too similar to recently sent ones.

The framework offers practical flexibility. Demotion logic can be customized for each dimension, and the strength of the penalty is tunable via weights per dimension (content, author, product type). That lets Instagram dial in the balance between personalization and variety rather than applying a one-size-fits-all rule.

How the Penalty Is Computed

Mechanically, each candidate’s final score is the product of its base ranking score and a penalty multiplier in the range [0,1]:

\text{Score}(c) = R(c) \times D(c)

For each dimension i, a similarity signal is computed between candidate c and the historical notification set H. The baseline implementation uses a binary signal—similarity is 1 if it exceeds a per-dimension threshold, else 0—based on a maximal marginal relevance (MMR) formulation:

p_i(c) = \mathrm{max}_{h \in H}\mathrm{sim}_i(c, h)

The penalties across dimensions are combined multiplicatively, with each weight wi in [0,1] controlling how strongly that dimension demotes similar candidates:

D(c) = \prod_{i=1}^{m} \left( 1 - w_i \cdot p_i(c) \right)

Because the penalty is multiplicative, a candidate similar to recent notifications along multiple dimensions gets demoted more heavily. High-relevance candidates that are genuinely fresh still pass through.

Results and What’s Next

The framework has already cut daily notification volume while improving CTR, suggesting users are getting fewer but more engaging alerts. Meta is also planning more adaptive demotion: rather than static rules, the strength of the penalty would respond to how many notifications a user has received and how quickly they arrive. If notifications are piling up or arriving in rapid succession, demotion would strengthen automatically.

Longer term, the team sees a role for large language models in the diversity pipeline. LLMs could assess semantic similarity between messages more deeply than surface-level rules and help rephrase notification content with greater variety. That would allow richer language and better relevance while still keeping the mix of topics, tone, and timing diverse.