- Meta shares how it applies machine learning (ML) and diversity algorithms to improve notification quality and user experience.
- A diverse notification ranking framework has been introduced to reduce uniformity and provide a more diverse and attractive notification combination.
- This new framework reduces the amount of notifications and drives higher engagement rates through more diverse outreach.
Notifications are one of the most powerful tools to bring people back to Instagram and enhance engagement. Whether it's a friend who likes your photos, another close friend who posts a story, or a reel suggestion you might enjoy, notifications can help you with important surface moments in real time.
Instagram leverages machine learning (ML) models to determine who gets notifications, when to send them, and what content to include. These models are trained to optimize positive user engagement, such as click-through rate (CTR) (the probability that a user will click a notification), and metrics such as time spent.
However, while an engagement-optimized model is effective in promoting interactions, there is a risk of overexpanding the types of products and authors that someone previously involved in. This can lead to excessive exposure to the same creator or the same product type, overlooking other valuable and diverse experiences.
This means that people can miss out on content that gives them a more balanced, satisfying, and richer experience. Over time, this will increase the likelihood that notifications will feel spam and people will disable them entirely.
The real challenge is finding the right balance. How can you introduce meaningful diversity into your notification experience without sacrificing Instagram personalization and relevance?
To tackle this, we have introduced a diversity-aware notification ranking framework that helps provide more diverse, better curation, and no-repeat notifications. This framework significantly reduces daily notification volume while improving CTR. We'll also introduce some of the benefits.
- It allows for scalability to incorporate customized soft penalty logic into each dimension, allowing for a more adaptive and sophisticated diversity strategy.
- Flexibility Tune demoted strength across dimensions such as content, author, product type, and more through adjustable weights.
- A balanced integration of personalization and diversity ensures notifications are both relevant and diverse.
Risk of non-diversity notifications
Overexposure issues in notifications are often presented in two main ways.
Overexposure to the same author: People may receive notifications that they are almost the same friends. For example, if someone frequently interacts with content from a particular friend, the system can continue to surface notifications from that person alone. This allows you to feel one-dimensional with repetition and reduces the overall value of the notification.
Overexposure to the same product surface: Even if feeds and reels can provide value, people may receive notifications primarily from the same product side, such as stories. For example, someone might be interested in notifications for both reels and storylines, but recently they've been interacting with stories more frequently. The system places a great priority on past engagement, so it sends only story notifications and looks down on the broader interests of a person.
Introducing Instagram's diversity response notification ranking framework
Instagram's Diversity-Aware Notification Ranking Framework is designed to enhance the notification experience by balancing the predictability of user engagement with the need for diversity in content. This framework introduces a diversity layer on top of existing engagement ML models and applies a multiplication penalty to the candidate scores generated by these models, as shown in Figure 1 below.
Diversity Layer evaluates the similarity of each candidate notification as a notification that has been recently sent across multiple dimensions, including content, author, notification type, and product surface. Next, we apply a carefully adjusted penalty (extracted as multiply demotion factor) to similar or repeating candidates. Adjusted scores are used to rerank candidates, allowing the system to select notifications that maintain high engagement potential while introducing meaningful diversity. Ultimately, the high quality bar will select top rank candidates that pass both ranking and diversity criteria.
Mathematical Formulation
Within the diversity class, apply a multiplier demotion factor to the base-related scores of each candidate. When a candidate notification is given, its final score is calculated as the product of the base ranking score and the diversity descent multiplier.
where R(c) Represents the candidate's base-related score d(c)∈ [0,1] is a penalty factor that reduces the score based on similarity to recently sent notifications. Define a set of semantic dimensions (authors, product types, etc.) that you want to promote diversity. For each dimension I, Calculate similarity signals pI(c) Between candidates c Set of history notifications huse a Maximum Relevance (MMR) approach.
where SimI(・・・) A predefined similarity function for dimension i. In the implementation of baseline, pI(c) Binary: Similarity equals 1 if the threshold is exceeded 𝜏I and 0 Otherwise.
The final demotion multiplier is defined as follows:
each wI ∈ [0,1] Controls the demolition strength of each dimension. This formulation ensures that candidates similar to previously delivered notifications along one or more dimensions are proportionally downward, reducing redundancy and promoting content variation. Multiplicative penalties allow flexible control across multiple dimensions to maintain high-related candidates.
The future of diversity-aware rankings
As we continue to evolve our ranking systems that recognize notification diversity, the next step is to introduce more adaptive and dynamic demolition strategies. Instead of relying on static rules, we plan to match demolition strength to the amount of notifications and timing of delivery. For example, as users receive more notifications, especially in similar types or rapid succession, the system gradually applies strong penalties to new notification candidates, effectively alleviating the overwhelming experience caused by high notification volumes or tight interval delivery.
In the long term, there is an opportunity to bring large-scale language models (LLMs) into the diversity pipeline. LLMS helps you go beyond surface-level rules by understanding semantic similarities between messages and resetting content in a more diverse and user-friendly way. This allows you to personalize and improve relevance of notification experiences in a richer language, while maintaining diversity across topics, tones and timings.
