Scorers

Scorers are responsible for assigning numerical scores to candidates in the Home Mixer pipeline. These scores determine the final ranking order of posts shown to users. Scorers implement the Scorer trait with both score and update methods.

Overview

Scorers generate new score values and return updated candidate objects. The update method merges these scores back into the original candidates:

#[async_trait]
trait Scorer<Q, C> {
    async fn score(&self, query: &Q, candidates: &[C]) -> Result<Vec<C>, String>;
    fn update(&self, candidate: &mut C, scored: C);
}

Available Scorers

PhoenixScorer

Uses the Phoenix ML prediction service to generate engagement probability scores for each candidate.

Arc<dyn PhoenixPredictionClient>

required

Client for communicating with the Phoenix prediction service

Purpose: Predicts user engagement likelihood across multiple action types

Predicted Actions

Phoenix generates probability scores for the following actions:

Option<f64>

Probability the user will like/favorite the tweet

Option<f64>

Probability the user will reply to the tweet

Option<f64>

Probability the user will retweet

Option<f64>

Probability the user will expand photos

Option<f64>

Probability the user will click the tweet

Option<f64>

Probability the user will click the author’s profile

Option<f64>

Probability of video quality view (watching video to completion)

Probability the user will share the tweet

Probability of sharing via direct message

Probability of sharing via copy link

Option<f64>

Probability the user will dwell on the tweet

Option<f64>

Probability the user will quote tweet

Option<f64>

Probability of clicking a quoted tweet

Option<f64>

Probability the user will follow the author

Option<f64>

Probability of negative feedback (“not interested”)

Option<f64>

Probability the user will block the author

Option<f64>

Probability the user will mute the author

Option<f64>

Probability the user will report the tweet

Continuous Actions

Option<f64>

Predicted dwell time in seconds (continuous value, not a probability)

Additional Metadata

Option<u64>

Unique identifier for the prediction request, useful for debugging and logging

Option<u64>

Timestamp (milliseconds since epoch) when the candidate was scored

Implementation Details

Code Example:

home-mixer/scorers/phoenix_scorer.rs

let tweet_infos: Vec<xai_recsys_proto::TweetInfo> = candidates
    .iter()
    .map(|c| {
        let tweet_id = c.retweeted_tweet_id.unwrap_or(c.tweet_id as u64);
        let author_id = c.retweeted_user_id.unwrap_or(c.author_id);
        xai_recsys_proto::TweetInfo {
            tweet_id,
            author_id,
            ..Default::default()
        }
    })
    .collect();

let result = self
    .phoenix_client
    .predict(user_id, sequence.clone(), tweet_infos)
    .await;

Retweet Handling: For retweets, Phoenix looks up predictions using the original tweet ID, ensuring consistent scoring:

home-mixer/scorers/phoenix_scorer.rs

let lookup_tweet_id = c.retweeted_tweet_id.unwrap_or(c.tweet_id as u64);

let phoenix_scores = predictions_map
    .get(&lookup_tweet_id)
    .map(|preds| self.extract_phoenix_scores(preds))
    .unwrap_or_default();

WeightedScorer

Combines Phoenix prediction scores into a single weighted score using configurable weights. Purpose: Translates multiple engagement signals into a single ranking score

Score Calculation

The weighted score is computed as:

home-mixer/scorers/weighted_scorer.rs

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);

Where apply multiplies each score by its weight:

fn apply(score: Option<f64>, weight: f64) -> f64 {
    score.unwrap_or(0.0) * weight
}

Configuration Weights

Weights are configured via constants (in the excluded params module):

Positive Engagement: FAVORITE_WEIGHT, REPLY_WEIGHT, RETWEET_WEIGHT, SHARE_WEIGHT, etc.
Negative Signals: NOT_INTERESTED_WEIGHT, BLOCK_AUTHOR_WEIGHT, MUTE_AUTHOR_WEIGHT, REPORT_WEIGHT (typically negative)
Continuous: CONT_DWELL_TIME_WEIGHT for dwell time prediction

Video Quality View (VQV) Eligibility

VQV weight is only applied to videos meeting minimum duration requirements:

home-mixer/scorers/weighted_scorer.rs

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
    }
}

Score Normalization

The weighted score is normalized and offset-adjusted:

home-mixer/scorers/weighted_scorer.rs

fn offset_score(combined_score: f64) -> f64 {
    if p::WEIGHTS_SUM == 0.0 {
        combined_score.max(0.0)
    } else if combined_score < 0.0 {
        (combined_score + p::NEGATIVE_WEIGHTS_SUM) / p::WEIGHTS_SUM * p::NEGATIVE_SCORES_OFFSET
    } else {
        combined_score + p::NEGATIVE_SCORES_OFFSET
    }
}

This ensures proper handling of negative signals while maintaining score scale.

Output

Option<f64>

The combined, weighted score used for initial ranking

AuthorDiversityScorer

Applies a decay multiplier to diversify authors within a feed response, preventing dominance by any single author.

f64

default:"AUTHOR_DIVERSITY_DECAY"

Exponential decay factor applied to subsequent posts from the same author

f64

default:"AUTHOR_DIVERSITY_FLOOR"

Minimum multiplier value, ensuring even repeated authors aren’t completely suppressed

Purpose: Increases author variety in the feed

Diversity Multiplier Formula

home-mixer/scorers/author_diversity_scorer.rs

fn multiplier(&self, position: usize) -> f64 {
    (1.0 - self.floor) * self.decay_factor.powf(position as f64) + self.floor
}

Where position is the 0-indexed count of previous posts from the same author. Example: If decay_factor = 0.8 and floor = 0.5:

1st post from author: multiplier = 1.0 (no penalty)
2nd post from author: multiplier = 0.9 (0.5 * 0.8^1 + 0.5)
3rd post from author: multiplier = 0.82 (0.5 * 0.8^2 + 0.5)
Converges to floor = 0.5

Algorithm

Sort candidates by weighted_score (descending)
Track author occurrence counts
For each author’s nth post, multiply score by multiplier(n-1)
Return updated scores in original order

Code Example:

home-mixer/scorers/author_diversity_scorer.rs

let mut author_counts: HashMap<u64, usize> = HashMap::new();

let mut ordered: Vec<(usize, &PostCandidate)> = candidates.iter().enumerate().collect();
ordered.sort_by(|(_, a), (_, b)| {
    let a_score = a.weighted_score.unwrap_or(f64::NEG_INFINITY);
    let b_score = b.weighted_score.unwrap_or(f64::NEG_INFINITY);
    b_score.partial_cmp(&a_score).unwrap_or(Ordering::Equal)
});

for (original_idx, candidate) in ordered {
    let entry = author_counts.entry(candidate.author_id).or_insert(0);
    let position = *entry;
    *entry += 1;

    let multiplier = self.multiplier(position);
    let adjusted_score = candidate.weighted_score.map(|score| score * multiplier);

    scored[original_idx] = PostCandidate {
        score: adjusted_score,
        ..Default::default()
    };
}

Output

Option<f64>

The diversity-adjusted score (weighted_score × diversity_multiplier)

OONScorer

Out-of-Network Scorer that applies a weight factor to out-of-network candidates, typically reducing their scores relative to in-network content.

f64

Multiplier applied to out-of-network candidates (configured via OON_WEIGHT_FACTOR)

Purpose: Balances in-network and out-of-network content in the feed Logic:

If candidate.in_network is Some(false), multiply score by OON_WEIGHT_FACTOR
Otherwise, leave score unchanged

Implementation:

home-mixer/scorers/oon_scorer.rs

let scored = candidates
    .iter()
    .map(|c| {
        let updated_score = c.score.map(|base_score| match c.in_network {
            Some(false) => base_score * p::OON_WEIGHT_FACTOR,
            _ => base_score,
        });

        PostCandidate {
            score: updated_score,
            ..Default::default()
        }
    })
    .collect();

This allows the system to prioritize familiar content (in-network) while still surfacing discovery content (out-of-network) when it’s highly relevant.

Output

Option<f64>

The network-adjusted final score

Scorer Pipeline Order

Scorers are typically applied in this sequence:

PhoenixScorer - Generate ML prediction scores
WeightedScorer - Combine predictions into weighted_score
AuthorDiversityScorer - Apply diversity adjustments to create score
OONScorer - Apply final network-based adjustments to score

Each scorer reads from previous scores and writes to new fields, creating a progression:

phoenix_scores → weighted_score → score (with diversity) → score (final)

Score Field Progression

Field	Type	Set By	Purpose
`phoenix_scores`	`PhoenixScores`	PhoenixScorer	Raw ML predictions
`weighted_score`	`Option<f64>`	WeightedScorer	Weighted combination
`score`	`Option<f64>`	AuthorDiversityScorer, OONScorer	Final ranking score

The final score field is used by selectors to determine the order candidates are served to users.

Sources - Retrieve candidates
Filters - Filter candidates before scoring
Selectors - Select and order scored candidates

Candidate Pipeline Traits

Phoenix Models

Home Mixer Components

Overview

Available Scorers

PhoenixScorer

Predicted Actions

Continuous Actions

Additional Metadata

Implementation Details

WeightedScorer

Score Calculation

Configuration Weights

Video Quality View (VQV) Eligibility

Score Normalization

Output

AuthorDiversityScorer

Diversity Multiplier Formula

Algorithm

Output

OONScorer

Output

Scorer Pipeline Order

Score Field Progression

​Overview

​Available Scorers

​PhoenixScorer

​Predicted Actions

​Continuous Actions

​Additional Metadata

​Implementation Details

​WeightedScorer

​Score Calculation

​Configuration Weights

​Video Quality View (VQV) Eligibility

​Score Normalization

​Output

​AuthorDiversityScorer

​Diversity Multiplier Formula

​Algorithm

​Output

​OONScorer

​Output

​Scorer Pipeline Order

​Score Field Progression

​Related Components

Overview

Available Scorers

PhoenixScorer

Predicted Actions

Continuous Actions

Additional Metadata

Implementation Details

WeightedScorer

Score Calculation

Configuration Weights

Video Quality View (VQV) Eligibility

Score Normalization

Output

AuthorDiversityScorer

Diversity Multiplier Formula

Algorithm

Output

OONScorer

Output

Scorer Pipeline Order

Score Field Progression

Related Components