Meta has implemented a new machine learning framework for Instagram that applies diversity algorithms to reduce repetitive content while maintaining user engagement. The diversity-aware ranking system addresses overexposure to similar content creators and product types by introducing multiplicative penalties to existing engagement models.
The framework tackles two primary problems: excessive messages from the same content creator and overemphasis on single-product surfaces like Stories while neglecting Feed or Reels content. Instagram’s machine learning models were previously optimized primarily for click-through rates and engagement metrics, which led to users receiving repetitive messages that could feel spammy and prompt disabling.
According to Instagram Engineers:
The real challenge lies in finding the right balance: How can we introduce meaningful diversity into the notification experience without sacrificing the personalization and relevance people on Instagram?
The new system operates as a diversity layer on top of existing engagement models. Notification candidates are evaluated across multiple dimensions, including content type, author identity, notification category, and product surface. For candidates deemed too similar to recent notifications, the framework applies calibrated multiplicative penalties that reduce their relevance score. A demotion multiplier, ranging from 0 to 1, adjusts the base score and lowers the rank of redundant notifications. Engineers can configure weights for each dimension to fine-tune the balance between relevance and diversity, giving different teams flexibility to adapt the framework to their product needs.
Instagram’s diversity-aware ranking framework(Source: Engineering at Meta Blog Post)
The mathematical implementation uses a base relevance score multiplied by a diversity demotion factor ranging from zero to one. For each semantic dimension, the system computes similarity signals between notification candidates and historical notifications using a maximal marginal relevance approach. Binary similarity indicators determine whether candidates exceed predefined thresholds for each dimension.
Instagram engineers report the framework has significantly reduced daily notification volume while improving click-through rates. The system provides extensibility for incorporating customized penalty logic across different dimensions and flexibility for adjusting demotion strength through configurable weights. The approach aimed to balance personalization with diversity to ensure notifications remain relevant while introducing variety.
According to the Instagram team, Future directions include exploring dynamic demotion strategies where penalty strength adapts to context, such as notification timing or frequency. As per the engineering team, it plans to investigate how large language models might measure semantic similarity to enhance notification diversity.
As mentioned by the Instagram and Meta engineers, the company’s approach reflects a broader trend in machine learning applications, where ranking systems are adapted to manage personalization and diversity. Similar techniques can be applied in recommendation systems, search engines, and ranking platforms to reduce redundancy while maintaining relevance.
