How the New-Author Boost Works in X-Algorithm: Cold-Start Detection and Scoring

The new-author boost is a temporary scoring multiplier that elevates posts from cold-start authors during the Home-Mixer ranking stage by applying a feature-switch controlled parameter through the AuthorRulesEvaluator.

The new-author boost mechanism in the x-algorithm repository addresses the cold-start problem by temporarily increasing the visibility of content from new authors. This implementation leverages the X-AI feature-switches system to dynamically adjust relevance scores while maintaining rigorous A/B testing standards through atomic impression tracking.

Feature-Switch Evaluation and Parameter Retrieval

The boost logic originates in home-mixer/util/author_rules.rs, where the AuthorRulesEvaluator interfaces with the feature-switch system to retrieve per-author experiment parameters.

The evaluator constructs a unique cache key—impress_cache_key—from the author ID and experiment name to ensure data integrity. When processing a candidate, the system queries for the rust_home_mixer_boost parameter, which defaults to 1.0 (neutral) but can be overridden by active experiments. This parameter retrieval mechanism allows product teams to deploy graduated boost values—such as 1.5 for treatment groups—without code deployments.

Cold-Start Detection and Score Multiplication

Actual score augmentation occurs in home-mixer/scorers/author_cold_start.rs, where the ColdStartScorer determines whether an author qualifies for the boost.

let boost = self.author_rules_evaluator.get(author_id, rust_home_mixer_boost);
if boost > 1.0 {
    // Apply the boost to the candidate's base score
    candidate_score *= boost;
}

When the retrieved multiplier exceeds 1.0, the scorer applies the factor to the candidate's base relevance score. This conditional multiplication ensures that only authors explicitly enrolled in boost experiments receive preferential treatment, while maintaining backward compatibility for standard ranking flows.

Integration with the Ranking Pipeline

The augmented scores flow into home-mixer/scorers/ranking_scorer.rs, where they integrate with the broader relevance calculation. The system applies the boost as specialized weights—including bidirectional_follow_reply_weight_boost and bidirectional_follow_dwell_weight_boost—that compound with standard engagement signals like dwell time and reply rates.

This architecture separates the boost detection logic from the final ranking arithmetic, allowing the ranking_scorer to treat the new-author multiplier as a first-class signal alongside traditional relevance features.

One-Time Impression Logging

The implementation guarantees measurement accuracy through atomic impression tracking. The AuthorRulesEvaluator utilizes a first_touch flag to ensure that boost experiment impressions are logged exactly once per author, preventing statistical contamination from repeated exposures.

This atomic logging mechanism enables reliable A/B test analysis while allowing the scoring system to reference the cached boost value throughout the session without re-triggering impression events.

Summary

• The new-author boost targets cold-start authors by assigning temporary score multipliers through the Home-Mixer pipeline.
• The AuthorRulesEvaluator in home-mixer/util/author_rules.rs retrieves the rust_home_mixer_boost parameter from feature-switch experiments, defaulting to 1.0.
• author_cold_start.rs applies the multiplier only when the boost value exceeds 1.0, multiplying the candidate's base score accordingly.
• The boost integrates into final rankings via specialized weight parameters in ranking_scorer.rs, combining with standard engagement signals.
• Atomic first_touch impression tracking ensures clean A/B test measurement by logging experiment participation exactly once per author.

Frequently Asked Questions

What threshold determines if an author receives the new-author boost?

The system applies the boost when the rust_home_mixer_boost parameter retrieved from the feature-switch system exceeds 1.0. The default value of 1.0 represents a neutral multiplier, while experiment treatments typically set values such as 1.5 to increase visibility.

How does x-algorithm prevent the same author from triggering multiple impression logs?

The AuthorRulesEvaluator implements an atomic first_touch flag within the impression caching logic. When constructing the cache key using impress_cache_key, the system checks this flag to ensure the experiment impression is recorded exactly once per author, regardless of how many times their content appears in subsequent ranking passes.

Where does the new-author boost fit into the overall ranking sequence?

The boost applies during the intermediate scoring stage in home-mixer/scorers/author_cold_start.rs before final ranking aggregation. After the ColdStartScorer calculates the multiplied score, the values propagate to home-mixer/scorers/ranking_scorer.rs where they merge with bidirectional follow weights and other relevance signals to produce the final content ordering.

Can the new-author boost value vary between different experiments?

Yes. The X-AI feature-switches system supports per-author experiment assignment, allowing different author cohorts to receive distinct boost values simultaneously. The AuthorRulesEvaluator handles these overrides independently, enabling multivariate testing of boost magnitudes without cross-contamination between experiment cells.

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