Twitter's recommendation system uses multiple interconnected models:

Core Ranking Models

Real-graph: Predicts interaction likelihood between users

• Determines if your followers will engage with your content
• Affects how widely Twitter shows your tweet to others
• Key signal: Will followers like, reply, or retweet this?

SimClusters: Community detection with sparse embeddings

• Identifies communities of users with similar interests
• Determines if your tweet resonates within specific communities
• Key strategy: Make content that appeals to tight communities who will engage

TwHIN: Knowledge graph embeddings for users and posts

• Maps relationships between users and content topics
• Helps Twitter understand if your tweet fits your follower interests
• Key strategy: Stay in your niche or clearly signal topic shifts

Tweepcred: User reputation/authority scoring

• Higher-credibility users get more distribution
• Your past engagement history affects current tweet reach
• Key strategy: Build reputation through consistent engagement

Engagement Signals Tracked

Twitter's Unified User Actions service tracks both explicit and implicit signals:

Explicit Signals (high weight):

• Likes (direct positive signal)
• Replies (indicates valuable content worth discussing)
• Retweets (strongest signal - users want to share it)
• Quote tweets (engaged discussion)

Implicit Signals (also weighted):

• Profile visits (curiosity about the author)
• Clicks/link clicks (content deemed useful enough to explore)
• Time spent (users reading/considering your tweet)
• Saves/bookmarks (plan to return later)

Negative Signals:

• Block/report (Twitter penalizes this heavily)
• Mute/unfollow (person doesn't want your content)
• Skip/scroll past quickly (low engagement)

The Feed Generation Process

Your tweet reaches users through this pipeline:

1. Candidate Retrieval - Multiple sources find candidate tweets:

   • Search Index (relevant keyword matches)
   • UTEG (timeline engagement graph - following relationships)
   • Tweet-mixer (trending/viral content)

2. Ranking - ML models rank candidates by predicted engagement:

   • Will THIS user engage with THIS tweet?
   • How quickly will engagement happen?
   • Will it spread to non-followers?

3. Filtering - Remove blocked content, apply preferences

4. Delivery - Show ranked feed to user

Optimization Strategies Based on Algorithm Insights

1. Maximize Real-graph (Follower Engagement)

Strategy: Make content your followers WILL engage with

• Know your audience: Reference topics they care about
• Ask questions: Direct questions get more replies than statements
• Create controversy (safely): Debate attracts engagement (but avoid blocks/reports)
• Tag related creators: Increases visibility through networks
• Post when followers are active: Better early engagement means better ranking

Example Optimization:

…

markdown-token-optimizer

分析并精简 Markdown 内容，降低 token 消耗并提升 AI 处理效率。