x-algo-scoring

# X Algorithm Scoring The X algorithm calculates a **weighted engagement score** for each post by combining predicted probabilities of 18 user actions. This score determines feed ranking. ## Weighted Score Formula ``` Score = Σ(weight × P(action)) for all 18 actions + offset ``` From `home-mixer/scorers/weighted_scorer.rs`: ```rust fn compute_weighted_score(candidate: &PostCandidate) -> f64 { let s: &PhoenixScores = &candidate.phoenix_scores; let vqv_weight = Self::vqv_weight_eligibility(candidate); let combined_score = Self::apply(s.favorite_score, p::FAVORITE_WEIGHT) + Self::apply(s.reply_score, p::REPLY_WEIGHT) + Self::apply(s.retweet_score, p::RETWEET_WEIGHT) + Self::apply(s.photo_expand_score, p::PHOTO_EXPAND_WEIGHT) + Self::apply(s.click_score, p::CLICK_WEIGHT) + Self::apply(s.profile_click_score, p::PROFILE_CLICK_WEIGHT) + Self::apply(s.vqv_score, vqv_weight) + Self::apply(s.share_score, p::SHARE_WEIGHT) + Self::apply(s.share_via_dm_score, p::SHARE_VIA_DM_WEIGHT) + Self::apply(s.share_via_copy_link_score, p::SHARE_VIA_COPY_LINK_WEIGHT) + Self::apply(s.dwell_score, p::DWELL_WEIGHT) + Self::apply(s.quote_score, p::QUOTE_WEIGHT) + Self::apply(s.quoted_click_score, p::QUOTED_CLICK_WEIGHT) + Self::apply(s.dwell_time, p::CONT_DWELL_TIME_WEIGHT) + Self::apply(s.follow_author_score, p::FOLLOW_AUTHOR_WEIGHT) + Self::apply(s.not_interested_score, p::NOT_INTERESTED_WEIGHT) + Self::apply(s.block_author_score, p::BLOCK_AUTHOR_WEIGHT) + Self::apply(s.mute_author_score, p::MUTE_AUTHOR_WEIGHT) + Self::apply(s.report_score, p::REPORT_WEIGHT); Self::offset_score(combined_score) } ``` ## Action Weights by Category ### Positive Weights (Increase Score) | Action | Weight Constant | Signal Type | | ------------------- | ---------------------------- | ------------------------------ | | Favorite | `FAVORITE_WEIGHT` | High value engagement | | Reply | `REPLY_WEIGHT` | High value engagement | | Retweet | `RETWEET_WEIGHT` | High value engagement | | Quote | `QUOTE_WEIGHT` | High value engagement | | Follow Author | `FOLLOW_AUTHOR_WEIGHT` | Very high value | | Share | `SHARE_WEIGHT` | Distribution signal | | Share via DM | `SHARE_VIA_DM_WEIGHT` | Distribution signal | | Share via Copy Link | `SHARE_VIA_COPY_LINK_WEIGHT` | Distribution signal | | Photo Expand | `PHOTO_EXPAND_WEIGHT` | Interest signal | | Click | `CLICK_WEIGHT` | Interest signal | | Profile Click | `PROFILE_CLICK_WEIGHT` | Interest signal | | VQV | `VQV_WEIGHT` | Video engagement (conditional) | | Dwell | `DWELL_WEIGHT` | Attention signal | | Quoted Click | `QUOTED_CLICK_WEIGHT` | Interest signal | | Dwell Time | `CONT_DWELL_TIME_WEIGHT` | Continuous attention | ### Negative Weights (Decrease Score) | Action | Weight Constant | Signal Type | | -------------- | ----------------------- | ------------------ | | Not Interested | `NOT_INTERESTED_WEIGHT` | Negative signal | | Block Author | `BLOCK_AUTHOR_WEIGHT` | Strong negative | | Mute Author | `MUTE_AUTHOR_WEIGHT` | Strong negative | | Report | `REPORT_WEIGHT` | Strongest negative | ## VQV Video Eligibility Video Quality View (VQV) weight only applies if video meets minimum duration: ```rust fn vqv_weight_eligibility(candidate: &PostCandidate) -> f64 { if candidate .video_duration_ms .is_some_and(|ms| ms > p::MIN_VIDEO_DURATION_MS) { p::VQV_WEIGHT } else { 0.0 // No VQV contribution for short videos or non-videos } } ``` ## Score Offset Logic Handles negative combined scores to ensure proper ranking: ```rust fn offset_score(combined_score: f64) -> f64 { if p::WEIGHTS_SUM == 0.0 { combined_score.max(0.0) } else if combined_score < 0.0 { // Negative scores get scaled offset (combined_score + p::NEGATIVE_WEIGHTS_SUM) / p::WEIGHTS_SUM * p::NEGATIVE_SCORES_OFFSET } else { // Positive scores just add offset combined_score + p::NEGATIVE_SCORES_OFFSET } } ``` ## Score Normalization After weighted scoring, scores are normalized (implementation in `util/score_normalizer.rs`, excluded from open source): ```rust let weighted_score = Self::compute_weighted_score(c); let normalized_weighted_score = normalize_score(c, weighted_score); ``` ## Additional Scoring Stages ### 1. Author Diversity Scoring Penalizes multiple posts from the same author to promote variety: ```rust // From home-mixer/scorers/author_diversity_scorer.rs fn multiplier(&self, position: usize) -> f64 { // First post from author: full score // Second post: score × decay_factor // Third post: score × decay_factor² (1.0 - self.floor) * self.decay_factor.powf(position as f64) + self.floor } ``` Parameters: `AUTHOR_DIVERSITY_DECAY`, `AUTHOR_DIVERSITY_FLOOR` ### 2. Out-of-Network Scoring Adjusts scores for posts from accounts user doesn't follow: ```rust // From home-mixer/scorers/oon_scorer.rs let updated_score = c.score.map(|base_score| match c.in_network { Some(false) => base_score * p::OON_WEIGHT_FACTOR, // Reduced weight _ => base_score, // Full weight for in-network }); ``` ## Example Score Calculation For a post with these predicted probabilities: - `favorite_score`: 0.12 (12% chance of like) - `reply_score`: 0.03 (3% chance of reply) - `retweet_score`: 0.05 (5% chance of retweet) - `not_interested_score`: 0.02 (2% chance of negative signal) ``` Weighted Score = 0.12 × FAVORITE_WEIGHT + 0.03 × REPLY_WEIGHT + 0.05 × RETWEET_WEIGHT + 0.02 × NOT_INTERESTED_WEIGHT (negative) + ... + offset ``` ## PostCandidate Score Fields ```rust pub struct PostCandidate { pub weighted_score: Option<f64>, // After WeightedScorer pub score: Option<f64>, // Final score after all scorers // ... } ``` ## Related Skills - `/x-algo-engagement` - Reference for all 18 action types - `/x-algo-pipeline` - Where scoring fits in the full pipeline