x-algo-engagement

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

X Algorithm Engagement Signals

X算法互动信号

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.

X推荐算法追踪18种互动行为类型以及1项连续指标。这些内容由Phoenix ML模型预测，用于计算加权分数。

PhoenixScores Struct

PhoenixScores 结构体

Defined in

home-mixer/candidate_pipeline/candidate.rs

rust

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>,
}

定义于

home-mixer/candidate_pipeline/candidate.rs

：

rust

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

按类别划分的行为类型

Positive Engagement (High Value)

正向互动（高价值）

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

行为	Proto 名称	描述
点赞	`ServerTweetFav`	用户点赞该帖子
回复	`ServerTweetReply`	用户回复该帖子
转发	`ServerTweetRetweet`	用户无评论转发帖子
引用转发	`ServerTweetQuote`	用户附带个人评论转发帖子
关注作者	`ClientTweetFollowAuthor`	用户关注该帖子的作者

Sharing Actions

分享行为

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

行为	Proto 名称	描述
分享	`ClientTweetShare`	通用分享行为
通过私信分享	`ClientTweetClickSendViaDirectMessage`	用户通过私信分享内容
通过复制链接分享	`ClientTweetShareViaCopyLink`	用户复制链接进行外部分享

Engagement Metrics

互动指标

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

行为	Proto 名称	描述
展开图片	`ClientTweetPhotoExpand`	用户展开图片查看
点击帖子	`ClientTweetClick`	用户点击该帖子
点击主页	`ClientTweetClickProfile`	用户点击作者的主页
VQV	`ClientTweetVideoQualityView`	视频质量观看（VQV）——用户观看视频达到有效时长
停留	`ClientTweetRecapDwelled`	用户在该帖子上停留（暂停浏览）
点击引用帖子	`ClientQuotedTweetClick`	用户点击引用的帖子

Negative Signals

负向信号

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

行为	Proto 名称	描述
不感兴趣	`ClientTweetNotInterestedIn`	用户标记为不感兴趣
屏蔽作者	`ClientTweetBlockAuthor`	用户屏蔽该作者
静音作者	`ClientTweetMuteAuthor`	用户静音该作者
举报	`ClientTweetReport`	用户举报该帖子

Continuous Actions

连续行为

Action	Proto Name	Description
Dwell Time	`DwellTime`	Continuous value: seconds spent viewing post

行为	Proto 名称	描述
停留时长	`DwellTime`	连续数值：用户查看帖子的秒数

How Scores Are Obtained

分数获取方式

The

PhoenixScorer

(

home-mixer/scorers/phoenix_scorer.rs

) calls the Phoenix prediction service:

Input: User history + candidate posts
Output: Log probabilities for each action type per candidate
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
    }
}

PhoenixScorer

（

home-mixer/scorers/phoenix_scorer.rs

）调用Phoenix预测服务：

输入：用户历史数据 + 候选帖子
输出：每个候选帖子对应每种行为类型的对数概率
转换：
```
概率 = exp(对数概率)
```

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),
        // ... 映射每个行为到对应的概率
    }
}

Signal Interpretation

信号解读

Scores are probabilities (0.0 to 1.0): P(user takes action | user sees post)
Higher = more likely: A
```
favorite_score
```
of 0.15 means 15% predicted chance of like
Negative signals have negative weights: High
```
report_score
```
reduces overall ranking
VQV requires minimum video duration: Only applies to videos >
```
MIN_VIDEO_DURATION_MS
```

分数为概率值（0.0 至 1.0）：P(用户看到帖子后执行该行为)
分数越高可能性越大：
```
favorite_score
```
为0.15意味着用户有15%的概率点赞该帖子
负向信号带有负权重：高
```
report_score
```
会降低帖子的整体排名
VQV要求最低视频时长：仅适用于时长超过
```
MIN_VIDEO_DURATION_MS
```
的视频

x-algo-engagement

Original

Translation

X Algorithm Engagement Signals

X算法互动信号

PhoenixScores Struct

PhoenixScores 结构体

Action Types by Category

按类别划分的行为类型

Positive Engagement (High Value)

正向互动（高价值）

Sharing Actions

分享行为

Engagement Metrics

互动指标

Negative Signals

负向信号

Continuous Actions

连续行为

How Scores Are Obtained

分数获取方式

Signal Interpretation

信号解读

Related Skills

相关技能