A New Diversity-Aware Ranking Framework for Better Instagram Notification Quality

TL;DR

• Meta is introducing a diversity-aware notification ranking framework for Instagram to balance personalization with content diversity and reduce repetitive notifications.
• The framework adds a diversity layer on top of existing engagement models to promote variety across authors, content types, and product surfaces while maintaining engagement potential.
• It uses multiplicative penalties and a set of dimensional similarity signals to downrank candidates that are too similar to recently sent notifications.
• Early results indicate a reduction in daily notification volume alongside improvements in click-through rate (CTR), signaling a better balance between relevance and variety.
• Looking ahead, the team plans adaptive, dynamic demotion strategies and, potentially, the involvement of large language models to enrich language and maintain diversity across topics and tone.

Context and background

Notifications are a powerful tool for reconnecting people with moments they might enjoy on Instagram. The platform relies on machine learning models to decide who should get a notification, when to deliver it, and what content to include. Those models have typically been optimized for user positive engagement metrics, such as click-through-rate (CTR) and time spent. However, maximizing engagement can risk overprioritizing the same authors or the same product types, leading to repetitive experiences that users may eventually find spammy. This can reduce overall satisfaction and even prompt users to disable notifications. To address this tension between personalization and diversity, Meta has developed a diversity-aware notification ranking framework. By introducing a diversity layer on top of the existing engagement models, the system aims to deliver more varied, better curated, and less repetitive notifications that still maintain meaningful engagement potential. The approach balances the predicted engagement with a broader set of experiences, reducing overexposure to any single author or product surface while expanding exposure to other relevant content. The claim is that this framework can reduce daily notification volume without sacrificing engagement, and in some cases even improve it Meta.

What’s new

The core innovation is a diversity layer that multiplies a candidate’s base relevance score by a diversity demotion multiplier. This layer evaluates each notification candidate against several semantic dimensions, such as author, content type, and product surface, to measure similarity with recently sent notifications. If a candidate is too similar to recent notices, it receives a downranked score due to the multiplicative penalties. The diversity signals are computed using a maximal marginal relevance approach for each dimension, producing a binary similarity signal p_i(c) that flags when similarity exceeds a threshold tau_i. The final score becomes the product of the base relevance score R(c) and a diversity demotion factor D(c) that lies in the range [0, 1]. Key design points include:

• Dimensions and signals: semantic dimensions like author and product type are used to promote diversity. For each dimension i, a similarity signal p_i(c) is computed with a similarity function sim_i(·, ·) against historical notifications H. In the baseline, p_i(c) is binary, equal to 1 if the similarity exceeds tau_i, and 0 otherwise.
• Demotion strength: each dimension has a weight w_i in [0, 1] that controls how strongly it demotes similar candidates. The final diversity multiplier D(c) reduces the base score when there is similarity across one or more dimensions.
• Scoring pipeline: the system computes a final score as the product of the candidate’s base relevance score and the diversity multiplier, enabling a re-ranked set that preserves engagement potential while injecting meaningful variety.
• Practical impact: the approach has shown reductions in daily notification volume while sustaining or improving CTR, suggesting users receive fewer notifications overall but with richer variety.
• Future directions: plans include more adaptive, dynamic demotion strategies that respond to notification volume and timing, and the potential inclusion of large language models to enrich language and broaden diversity across topics, tone, and timing.

Why it matters (impact for developers/enterprises)

For developers and product teams, this framework offers a path to balance two competing goals: delivering timely, relevant notifications and ensuring users are exposed to a broader range of content and authors. By downweighting overly repetitive candidates, the system can reduce user fatigue and the perceived spamminess of notifications, which is critical for user retention and long-term engagement. At the same time, preserving a high engagement potential means developers can maintain strong performance metrics such as CTR and time spent, while offering a more varied and meaningful notification experience. The approach also demonstrates how a modular diversity layer can be integrated with existing ML-based ranking systems. It provides a blueprint for other platforms seeking to introduce diversity controls without discarding proven relevance signals. While current focus is on Instagram notifications, the underlying concept – aggregating similarity signals across multiple dimensions and applying multiplicative penalties – could inform notification strategies across different surfaces and notification types on the platform.

Technical details or Implementation (overview)

• Base ranking: Each candidate c has a base relevance score R(c) generated by the existing engagement ML models that optimize for outcomes like CTR and time spent.
• Diversity layer inputs: A set of semantic dimensions i is defined (for example, author, content type, and product surface). For each dimension, a similarity signal p_i(c) is computed between candidate c and the historical notification set H using a dimension-specific similarity function sim_i(·, ·).
• Similarity signals: In the baseline, p_i(c) is binary and equals 1 if the similarity exceeds a threshold tau_i, and 0 otherwise.
• Demotion weights: Each dimension has a weight w_i in the interval [0, 1] that controls how strongly similarity reduces candidate scores.
• Final score computation: The final score for a candidate is the product of its base relevance score and a diversity demotion multiplier D(c) in [0, 1]. The candidate selection process then picks the notification with the highest final score that also passes the diversity criteria.
• Practical deployment: The diversity layer is designed to operate on top of the existing ranking framework, enabling a controlled introduction of diversity without abandoning relevance.
• Adaptive future work: As notification volume and delivery timing vary, stronger penalties may be applied dynamically to prevent overwhelming experiences, especially when similar notifications are delivered in rapid succession. There is also interest in exploring how large language models could help rephrase content and broaden semantic diversity while maintaining relevance and tone consistency.

Key takeaways

• A diversity-aware ranking framework adds a dedicated layer to promote variety in notifications while preserving engagement potential.
• The framework uses multiple semantic dimensions and binary similarity signals to downweight repetitive candidates.
• Final scoring combines the base relevance with a diversity multiplier to re-rank candidates for delivery.
• Early results indicate reductions in daily notification volume with maintained or improved CTR, suggesting a better balance between relevance and diversity.
• Future work includes adaptive demotion strategies and the potential use of LLMs to enrich language and semantic variety.

References

• Meta engineering blog - A new ranking framework for better notification quality on Instagram