How Instagram Reels Uses Recommender Systems and Machine Learning Models

Instagram Reels employs a multi-stage recommender system powered by machine learning models, primarily Two Towers Neural Networks, to deliver personalized short video recommendations based on user engagement and behavior.

Core Components of the Recommender System

The system analyzes extensive user engagement data, including viewing duration, likes, comments, shares, and saves, to predict preferences and curate feeds from billions of videos. It addresses challenges like the cold-start problem for new users by evaluating connections from followed accounts and their networks.

Key processes include:

• Candidate Retrieval: Pulls potential videos using interaction history and content similarity, filtering out low-quality items.
• Multi-Stage Ranking:

  Stage	Description	Models Used
  First-Stage	Lightweight ranking of thousands of candidates by predicted engagement (e.g., likes, shares).	Two Towers Neural Networks (caches data for efficiency).
  Second-Stage	Deeper analysis of user-item interactions to refine top candidates.	Heavy, complex models.
  Final Reranking	Evaluates top items (e.g., from 500 to 50) for relevance and quality.	Distillation models.

This pipeline enables real-time adaptation to evolving interests, trends, and content, preventing repetitive feeds.

Machine Learning Models and Scaling

Instagram uses advanced neural networks like Two Towers for efficient real-time processing of vast data volumes. The system has scaled to over 1000 ML models, handling high-scale retrieval and ranking without performance loss. Content is prioritized by relevance over creator popularity, using "seed" accounts for initial sampling.

User Controls and Personalization

Recent updates include the "Your Algorithm" tool (rolled out in the US by December 2025, expanding globally), allowing users to view AI-summarized topics (e.g., "creativity, sports hype") from their activity and adjust recommendations by selecting or adding interests. Each app section (Feed, Explore, Reels) has dedicated ranking algorithms.

These elements boost engagement by aligning content with individual tastes, fostering longer sessions. Sources note Instagram's system evolves rapidly, with potential for further expansions like Explore personalization.