用户被通知烦到关掉Instagram？Meta：我们反思了，用AI给自己“限流”

Core Insights - Meta has introduced a new machine learning ranking framework for Instagram that incorporates "diversity algorithms" to reduce content redundancy and alleviate user notification fatigue while maintaining overall engagement [1][5] - The new system addresses two main issues: excessive notifications from the same creator and the algorithm's bias towards a single content type, such as Stories, neglecting other formats like Feed or Reels [1][4] Summary by Sections New Framework Implementation - The new system operates as a "diversity filtering layer" on top of the existing interaction model, evaluating candidate notifications across multiple dimensions, including content type, author identity, notification category, and product area [2][4] - A calibrated "multiplicative penalty" is applied to candidates that are too similar to recent notifications, with a penalty coefficient ranging from 0 to 1, allowing for adjustments to the base score to lower the ranking of duplicate notifications [2][4] Mathematical Approach - The final score is calculated using the formula "base relevance score × diversity penalty factor," employing the "maximal marginal relevance" method to assess similarity signals between notification candidates and historical notifications [4] - This framework has significantly reduced the number of daily notifications received by users while enhancing click-through rates, with customizable penalty logic and adjustable weights for different dimensions to balance relevance and diversity [4] Future Developments - Instagram's team plans to explore "dynamic demotion strategies," which would allow penalty intensity to adjust automatically based on context, such as the timing or frequency of notifications [4] - There are intentions to investigate the use of large language models to measure semantic similarity for further optimization of notification diversity [4][5] Industry Trends - This approach reflects a broader trend in machine learning applications, where ranking systems are shifting from solely pursuing personalization to finding a balance between relevance and diversity, applicable in recommendation systems, search engines, and other ranking platforms [5]