The X recommendation algorithm tracks 18 engagement action types plus 1 continuous metric. These are predicted by the Phoenix ML model and used to calculate weighted scores.
PhoenixScores Struct
Defined in home-mixer/candidate_pipeline/candidate.rs:
pub struct PhoenixScores {
// Positive engagement signals
pub favorite_score: Option<f64>,
pub reply_score: Option<f64>,
pub retweet_score: Option<f64>,
pub quote_score: Option<f64>,
pub share_score: Option<f64>,
pub share_via_dm_score: Option<f64>,
pub share_via_copy_link_score: Option<f64>,
pub follow_author_score: Option<f64>,
// Engagement metrics
pub photo_expand_score: Option<f64>,
pub click_score: Option<f64>,
pub profile_click_score: Option<f64>,
pub vqv_score: Option<f64>, // Video Quality View
pub dwell_score: Option<f64>,
pub quoted_click_score: Option<f64>,
// Negative signals
pub not_interested_score: Option<f64>,
pub block_author_score: Option<f64>,
pub mute_author_score: Option<f64>,
pub report_score: Option<f64>,
// Continuous actions
pub dwell_time: Option<f64>,
}
## Action Types by Category
### Positive Engagement (High Value)
| Action | Proto Name | Description |
| ----------------- | ------------------------- | ----------------------------------- |
| **Favorite** | `ServerTweetFav` | User likes the post |
| **Reply** | `ServerTweetReply` | User replies to the post |
| **Retweet** | `ServerTweetRetweet` | User reposts without comment |
| **Quote** | `ServerTweetQuote` | User reposts with their own comment |
| **Follow Author** | `ClientTweetFollowAuthor` | User follows the post's author |
### Sharing Actions
| Action | Proto Name | Description |
| ----------------------- | -------------------------------------- | ------------------------------------ |
| **Share** | `ClientTweetShare` | Generic share action |
| **Share via DM** | `ClientTweetClickSendViaDirectMessage` | User shares via direct message |
| **Share via Copy Link** | `ClientTweetShareViaCopyLink` | User copies link to share externally |
### Engagement Metrics
| Action | Proto Name | Description |
| ----------------- | ----------------------------- | --------------------------------------------------------------- |
| **Photo Expand** | `ClientTweetPhotoExpand` | User expands photo to view |
| **Click** | `ClientTweetClick` | User clicks on the post |
| **Profile Click** | `ClientTweetClickProfile` | User clicks author's profile |
| **VQV** | `ClientTweetVideoQualityView` | Video Quality View - user watches video for meaningful duration |
| **Dwell** | `ClientTweetRecapDwelled` | User dwells (pauses) on the post |
| **Quoted Click** | `ClientQuotedTweetClick` | User clicks on a quoted post |
### Negative Signals
| Action | Proto Name | Description |
| ------------------ | ---------------------------- | ---------------------------- |
| **Not Interested** | `ClientTweetNotInterestedIn` | User marks as not interested |
| **Block Author** | `ClientTweetBlockAuthor` | User blocks the author |
| **Mute Author** | `ClientTweetMuteAuthor` | User mutes the author |
| **Report** | `ClientTweetReport` | User reports the post |
### Continuous Actions
| Action | Proto Name | Description |
| -------------- | ----------- | -------------------------------------------- |
| **Dwell Time** | `DwellTime` | Continuous value: seconds spent viewing post |
## How Scores Are Obtained
The `PhoenixScorer` (`home-mixer/scorers/phoenix_scorer.rs`) calls the Phoenix prediction service:
1. **Input**: User history + candidate posts
2. **Output**: Log probabilities for each action type per candidate
3. **Conversion**: `probability = exp(log_prob)`
```rust
fn extract_phoenix_scores(&self, p: &ActionPredictions) -> PhoenixScores {
PhoenixScores {
favorite_score: p.get(ActionName::ServerTweetFav),
reply_score: p.get(ActionName::ServerTweetReply),
retweet_score: p.get(ActionName::ServerTweetRetweet),
// ... maps each action to its probability
}
}
Signal Interpretation
- Scores are probabilities (0.0 to 1.0): P(user takes action | user sees post)
- Higher = more likely: A
favorite_scoreof 0.15 means 15% predicted chance of like - Negative signals have negative weights: High
report_scorereduces overall ranking - VQV requires minimum video duration: Only applies to videos >
MIN_VIDEO_DURATION_MS
Related Skills
/x-algo-scoring- How these signals are combined into a weighted score/x-algo-ml- How Phoenix model predicts these probabilities