How Instagram Dominates Using Explore Feed Ranking Systems

Introduction to Unfamiliar Account Recommendation

The Instagram Explore feed represents a distinct recommendation challenge compared to standard social feeds. While the Home feed primarily ranks content from accounts a user explicitly follows, the Explore feed must select content from millions of unfamiliar creators across the platform.

The system must analyze a user's historical actions to identify interests and recommend relevant content from a massive pool of daily uploads. To solve this, Instagram developed an account-centric candidate generation pipeline powered by custom embedding models and a multi-stage ranking framework.

Account Representation and Embedding Space

Instead of analyzing billions of individual posts directly, Instagram's system focuses on accounts. The architecture assumes that accounts producing cohesive content (e.g., travel photography, baking, software engineering) can be grouped in a dense vector space.

To map these relationships, Instagram developed Ig2Vec, a custom embedding model based on word2vec skip-gram concepts. Instead of analyzing word sequences in sentences, Ig2Vec treats the sequence of accounts a user interacts with in a single session as a "sentence." Accounts that are frequently interacted with together are placed close to one another in the 256-dimensional embedding space, establishing a dense representation of related interests.

This embedding framework is trained using Skip-Gram with Negative Sampling (SGNS) on historical interaction paths, allowing the system to identify a user's interest profile by evaluating their "seed accounts"—accounts they have recently liked, saved, or interacted with. The dense vector representations capture subtle affinities that simple tag matching or text analysis would miss entirely. Interaction weights are applied here, where a direct share or a post save is weighted significantly higher than a simple double-tap like or passive impression, ensuring the seed accounts reflect active user interest.

Seed-Based Retrieval, Filtering, and Content Moderation

The candidate generation phase begins by retrieving these seed accounts. The retrieval service queries an approximate nearest neighbor (ANN) index to find the top 500 accounts closest to the seeds in the Ig2Vec embedding space.

From these accounts, the system pulls a pool of media candidates (photos, Reels, and carousels) that are popular, high-quality, and recent. This reduces the initial pool of millions of media items to a manageable set of roughly 10,000 candidates. Content moderation filters are also applied at this stage, pruning sensitive or low-quality posts before they enter the heavy ranking pipelines.

Multi-Stage Funnel, Dual-Pass Ranking, and Blending

To rank this candidate pool with sub-100ms latency, the system routes the items through a multi-stage filtering funnel. The first stage utilizes a lightweight ranker, such as a logistic regression model or a small neural network, to evaluate basic features.

This stage quickly filters the 10,000 candidates down to 500. The second stage employs a deep neural network that evaluates rich user and media features. This heavy model predicts the probability of specific high-value user actions, such as saving the post, liking it, or sharing it with friends.

The final ranking score is a weighted combination of these predicted probabilities. By dividing the workload into a two-pass ranking topology, the system minimizes the processing required for irrelevant candidates, keeping hardware costs manageable.

• First-Pass Filtering: Evaluates coarse features via a lightweight model to filter candidates from 10,000 to 500.
• Second-Pass Deep Model: Analyzes dense user-media embeddings to predict specific engagement action probabilities.
• Weighted Score Aggregation: Combines predicted probabilities of likes, saves, and shares into a single ranking metric.

The final phase is re-ranking and blending, which applies safety, variety, and business rules to the top-scoring candidates. Integrity filters automatically remove content that violates community guidelines, while diversity rules ensure the feed is not dominated by a single creator or media format.

For example, if the top 10 items are all video Reels, the blending algorithm mixes in photos and carousels to maintain visual variety. The resulting balanced feed is then structured and returned to the client application.

Operating this multi-stage funnel at scale requires a highly optimized infrastructure. Instagram distributes its vector indices across a cluster of search nodes using custom sharding topologies.

To keep latency low, the system pre-computes candidate sets asynchronously for active users, caching these sets in memory. When a user refreshes their Explore page, the system retrieves the cached candidate set, applies the final ranking models, and serves the results, minimizing real-time computational load.

Technical Implementation: Cosine Similarity Vector Calculations

Calculating cosine similarity computes the directional alignment between high-dimensional vector representations, measuring the absolute distance within the account embedding space.

Account-Centric Discovery Engines at the Edge with Bramsley