How the Out-of-Network Discount Works in the X Feed Algorithm

The X feed algorithm applies a multiplicative discount of 25% (using a factor of 0.75) to out-of-network candidates via the configurable rust_home_mixer_oon_weight_factor parameter.

The out-of-network discount is a core mechanism in the X (formerly Twitter) feed ranking system that reduces the visibility of content originating from outside a viewer's social graph. According to the xai-org/x-algorithm source code, this discount ensures that posts from accounts the user does not follow receive lower base relevance scores than in-network content. The implementation relies on a configurable weight factor applied during the final scoring phase of the home feed mixer.

Understanding the Out-of-Network Discount Logic

When the RankingScorer processes candidates for the home timeline, it distinguishes between content from followed accounts (in-network) and content from accounts the user does not follow (out-of-network). For out-of-network candidates, the algorithm multiplies the base relevance score by the OON weight factor to de-prioritize external content.

The default configuration sets this factor to 0.75, meaning out-of-network posts receive only 75% of the score they would otherwise earn. This 25% reduction applies after all other relevance calculations but before final ranking, creating a tunable boundary between social graph content and external recommendations.

Implementation in the Ranking Scorer

The discount logic lives in home-mixer/scorers/ranking_scorer.rs, where the scorer retrieves the configuration flag and conditionally applies the multiplier.

Parameter Configuration

The weight factor is defined as a configurable parameter in home-mixer/params/param.rs. The system exposes this value through feature flags using the parameter name rust_home_mixer_oon_weight_factor, allowing real-time adjustment without code deployment. This configuration-based approach enables the platform to tune the boundary between in-network and out-of-network content dynamically.

Score Calculation Logic

Inside the RankingScorer implementation, the code retrieves the factor from query parameters and applies it during candidate evaluation:

// Retrieve the OON weight factor from feature flags
let oon_weight_factor = query.params.get(OonWeightFactor);

// Apply discount for out-of-network candidates
if !candidate.is_in_network {
    score = base_score * oon_weight_factor;
}

This logic executes for every candidate in the scoring pipeline. The is_in_network boolean property determines whether the candidate originates from the viewer's social graph, triggering the multiplication when the value is false.

Code Example: Verifying the Discount Behavior

The repository includes unit tests in home-mixer/scorers/ranking_scorer.rs that validate the discount calculation. The following pattern demonstrates how to verify that out-of-network candidates receive exactly 75% of their in-network equivalent scores:

// Configure the scorer with the default OON weight factor
let query = query_with_flags(&[
    ("rust_home_mixer_oon_weight_factor", "0.75"), // 25% discount
    ("rust_home_mixer_value_model_mode", "weighted"),
]);

// Create test candidates: one in-network, one out-of-network
let candidates = vec![
    candidate(1, Some(true)),   // in-network
    candidate(2, Some(false)),  // out-of-network
];

let scored = scorer.score(&query, &candidates).await;

// Verify the discount is applied correctly
let in_network_score = scored[0].as_ref().unwrap().score.unwrap();
let oon_score = scored[1].as_ref().unwrap().score.unwrap();
assert!((oon_score - in_network_score * 0.75).abs() < 1e-9);

This test confirms that when two candidates have identical base relevance scores, the out-of-network candidate's final score equals the in-network score multiplied by 0.75.

Configuration and Tuning

The out-of-network discount operates as a feature flag-controlled parameter, making it adjustable per request or through gradual rollouts. Platform operators can modify the rust_home_mixer_oon_weight_factor value to:

• Increase the discount (lower values) to prioritize social graph content more aggressively
• Decrease the discount (higher values) to surface more external recommendations
• Set the factor to 1.0 to effectively disable the out-of-network penalty

According to the source code analysis, the parameter retrieval happens at the beginning of the scoring process, ensuring the factor remains consistent across all candidates within a single ranking request.

Summary

• The out-of-network discount applies a multiplicative factor of 0.75 (default) to candidates outside the viewer's social graph.
• Implementation resides in home-mixer/scorers/ranking_scorer.rs within the RankingScorer logic.
• The configurable parameter rust_home_mixer_oon_weight_factor controls the discount magnitude via feature flags.
• Out-of-network candidates receive 75% of their base relevance score compared to identical in-network content.
• Unit tests verify the precise mathematical application of this discount in the scoring pipeline.

Frequently Asked Questions

What is the default out-of-network discount value in the X algorithm?

The default value is 0.75, which applies a 25% reduction to the base relevance scores of out-of-network candidates. This means content from accounts the user does not follow starts with a 25% penalty compared to social graph content before other ranking factors are considered.

How is the out-of-network discount applied in the X feed algorithm?

The discount applies as a multiplicative factor during the final scoring phase. When candidate.is_in_network evaluates to false, the RankingScorer multiplies the candidate's base score by the rust_home_mixer_oon_weight_factor. This occurs in home-mixer/scorers/ranking_scorer.rs after relevance calculations but before final ranking.

Can the out-of-network discount factor be adjusted without deploying new code?

Yes. The factor is configured through the rust_home_mixer_oon_weight_factor feature flag, defined in home-mixer/params/param.rs. This allows real-time adjustment of the discount rate via feature flag systems without requiring code changes or service restarts.

Where is the out-of-network discount logic implemented in the x-algorithm codebase?

The core logic resides in home-mixer/scorers/ranking_scorer.rs, specifically within the scoring method that checks the is_in_network property. The parameter definition exists in home-mixer/params/param.rs, and comprehensive unit tests verifying the 0.75 multiplier behavior appear in the test section of the ranking scorer file.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}