TechPrufer
Get started
Meta
Meta logoMeta · Feature trail

Social Recommendations System

1 post · Sep 2011 – Mar 2026

What this trail covers

Meta's social recommendations system has evolved to automatically extract and display place recommendations from user comments on posts. This involves natural language understanding to identify place mentions, entity resolution to match them to specific businesses, and dynamic feed updates across platforms to present this information seamlessly. The system aims to enhance local discovery and user engagement by leveraging social interactions. The Place Tips platform has been extended to support e-commerce recommendations, and the platform has scaled to manage over 1000 ML models for various ranking surfaces (Feed, Stories, Reels, comments, notifications, tagging), implementing a model registry, automated launch tooling, and a model stability metric to ensure reliability and accelerate ML velocity.

All Meta trails
  1. Friend Bubbles: Enhancing Social Discovery on Facebook ReelsThis post details the technical architecture of Friend Bubbles in Facebook Reels, an extension of the social recommendations system. It introduces two complementary ML models for Viewer-Friend Closeness: one based on user surveys and social graph features, and another context-specific model trained on on-platform activity signals (likes, comments, reshares). The system expands the top of the funnel by explicitly retrieving friend-interacted content. Ranking models are enhanced with friend-bubble interaction signals and new tasks to incorporate viewer-friend relationship strength and learn downstream engagement. A key metric used is P(video engagement | bubble impression). Client-side optimizations focus on smooth scrolling, no load latency regressions, and low CPU overhead by integrating metadata retrieval into the video prefetch window and conditionally disabling animations. The system aims to maximize overall video engagement by balancing social interaction and video engagement through tunable weights.
  2. A New Ranking Framework for Better Notification Quality on InstagramIntroduced a diversity-aware notification ranking framework that sits on top of existing engagement ML models. This framework applies multiplicative penalties to candidate notification scores based on similarity to recently sent notifications across dimensions like content, author, and product type. The mathematical formulation involves a base relevance score R(c) multiplied by a diversity demotion multiplier D(c), where D(c) is a product of (1 - w_i * p_i(c)) for each dimension i. p_i(c) is a similarity signal computed using maximal marginal relevance (MMR) against historical notifications H. Future work includes dynamic demotion strategies and LLM integration for semantic similarity.
  3. Journey to 1000 models: Scaling Instagram’s recommendation systemThis post details Instagram's scaling of its recommendation system to manage over 1000 ML models. Key contributions include the implementation of a model registry for tracking model importance and business function, automated model launch tooling that reduced launch time from days to hours, and the definition and operationalization of 'model stability' as a metric for measuring prediction accuracy and setting SLOs. The model registry standardizes metadata like 'model type' (e.g., 'ig_stories_tray_mtml') and 'criticality' (TIER0-TIER4), integrating with Meta's Configerator and SLICK for automated monitoring and alerting. The post also describes the ranking funnel (sourcing, ESR, LSR) and the challenges of managing diverse models across these layers.
  4. Made in NY: The engineering behind social recommendationsThis post details the engineering behind Meta's new Recommendations feature, which automatically extracts place recommendations from user comments. It covers the challenges of natural language understanding for recommendation extraction, entity resolution to map free-text mentions to specific businesses, and the development of dynamically updating feed attachments for both posts and comments. The technical stack for local search involves entity retrieval and scoring using models like LambdaMART, while dynamic feed rendering on Android and iOS utilizes ComponentKit and custom state management to ensure consistency across platforms.
  5. Under the hood: Place Tips at LollapaloozaThis post details the technical implementation of Place Tips for Lollapalooza, focusing on the server-controlled native client rendering framework. It introduces the concept of 'styles' as contracts between client and server for rendering and behavior, and demonstrates how abstractly composed sub-templates (header, attachments, footer) allow for flexible data aggregation and consistent UI across diverse app versions. The Lollapalooza implementation leveraged existing client styles and server-side logic for rapid development.
  7. Under the Hood: Building the App Center recommendation engineThis post details the initial development of the App Center recommendation engine, which uses an aggregator-leaf architecture for sharding and replication to handle large-scale data and traffic. It introduces a system for measuring app quality based on user ratings and daily active users, incorporating confidence adjustments. The algorithmic elements include candidate selection heuristics (demographics, social, past interactions), scoring based on explicit, dynamic, and learned latent features (matrix factorization), and real-time updates for indexing, user history, dynamic features, and latent features.
  8. Interning at Facebook: Who Goes Where When (and Why It Matters)This post details the development of an algorithm to predict business operating hours based on user check-in data. The algorithm involves making initial predictions based on an openness threshold, filling short gaps in predictions, and aggregating predictions across days. It also describes the implementation of user interfaces for suggesting and validating business hours, and the development of a 'check-in similarity score' to trust user suggestions. Finally, it outlines improvements to the user experience by displaying estimated and accepted hours on Place Pages and city Pages, leading to a significant increase in places with listed hours.