Instagram

Users annoyed into turning off Instagram notifications? Meta: We’ve reflected — using AI to limit ourselves

Honghao Wang

— 3 min read
Users annoyed into turning off Instagram notifications? Meta: We’ve reflected — using AI to limit ourselves

Instagram Introduces “Diversity Algorithms” in New Machine Learning Framework to Boost Engagement

image
image

Meta has rolled out a new machine learning ranking framework for Instagram, aiming to reduce repetitive notifications through diversity algorithms. This upgrade seeks to combat notification fatigue, ensure content variety, and maintain high engagement levels.

The system addresses the problem of overexposing users to similar creators or products by applying multiplicative penalties within Instagram’s existing engagement models.

---

Key Challenges Tackled

The framework focuses on two persistent problems:

  • Over-notification from the same creator
  • Heavy bias toward a single content type
  • Example: Stories dominating over Feed posts or Reels

Previously, models optimized solely for click-through and engagement metrics—effective for short-term activity but prone to spamming users with similar content, prompting them to disable notifications.

---

How the Diversity Layer Works

Operating as a “filtering layer” above the engagement model, this system evaluates notifications against multiple dimensions:

  • Content type
  • Author identity
  • Notification category
  • Product surface (Feed, Reels, Stories)

If a candidate notification is too similar to recent ones, the system applies a penalty factor (0 to 1) to reduce its relevance score.

image

Instagram's diversity ranking framework (source: Engineering at Meta blog)

Formula

Final Score = Base Relevance Score × Diversity Penalty Factor

Key method:

  • Uses Maximal Marginal Relevance (MMR) to detect similarity signals
  • Flags candidates as “similar” when exceeding thresholds in any dimension
  • Demotes flagged items in the ranking

Customizability: Different weights and thresholds can be set per dimension, allowing a tailored balance between relevance and diversity.

---

Results & Future Roadmap

Observed impacts:

  • Reduced average daily notifications per user
  • Increased click-through rates
  • Enhanced scalability for different product strategies

Next steps:

  • Dynamic demotion strategies: Penalty strength adapts to context, such as notification frequency or timing
  • LLM-based similarity metrics: Use large language models to improve semantic diversity detection

---

Broader Significance

This approach reflects a shift in ML-driven content surfacing—from pure personalization toward balanced relevance and diversity.

Potential applications include:

  • Recommendation engines
  • Search ranking systems
  • Any platform seeking to reduce redundancy and keep user experience fresh

---

Original source:

https://www.infoq.com/news/2025/09/instagram-notification-ranking/

---

Related: AI Tools for Content Creators

In parallel to algorithm evolution, creator tools are crucial.

Platforms like AiToEarn官网 provide an open-source ecosystem enabling AI-assisted content generation, cross-platform publishing, analytics, and monetization.

Supported networks include:

  • Douyin, Kwai, WeChat, Bilibili, Xiaohongshu (Rednote)
  • Facebook, Instagram, LinkedIn, Threads
  • YouTube, Pinterest, X (Twitter)

This helps creators distribute diverse content efficiently, avoid redundancy, and maximize reach.

---

image

---

Read the full article

Open in WeChat

---

Why This Matters for You

The rapid innovation in ranking algorithms and diversity strategies mirrors the challenges faced by digital creators, developers, and analysts.

By using tools like AiToEarn, you can:

  • Harness AI-driven creativity
  • Publish simultaneously across major social networks
  • Analyze engagement metrics
  • Optimize for both personalization and content variety

If you’d like, I can create a concise visual flow diagram showing how Instagram’s diversity algorithm processes notifications — would you like me to prepare that next?

Read more