🎯 How LinkedIn Builds Feed Ranking System
1️⃣ Core Feed Framework (Staff-Level)
When discussing a LinkedIn-like feed ranking system, I frame it as:
- Candidate generation
- Graph and follow relationships
- Feature enrichment
- Ranking model
- Quality and integrity filters
- Diversity and freshness controls
- Feedback loops
- Trade-offs: engagement vs professional value vs trust
2️⃣ Core Problem
LinkedIn feed is not pure entertainment ranking.
It should optimize:
- professional relevance
- network value
- content quality
- creator and company relevance
- freshness
- trust and safety
- long-term user value
👉 Interview Answer
A LinkedIn-like feed system ranks professional content, not just viral content. It must balance engagement with content quality, network relevance, trust, and long-term professional value.
3️⃣ High-Level Architecture
User opens feed
↓
Candidate Generation
↓
Feature Enrichment
↓
Ranking Model
↓
Quality / Spam Filters
↓
Diversity and Freshness Reranking
↓
Feed Response
↓
Impression and Engagement Feedback
4️⃣ Candidate Generation
Sources:
- first-degree connections
- followed companies
- followed creators
- groups
- liked or commented posts from network
- recommended professional content
- ads or sponsored content
Goal:
retrieve enough relevant candidates without doing expensive ranking over everything.
5️⃣ Ranking Signals
User features:
- industry
- role
- seniority
- skills
- network graph
- past engagement
Post features:
- author quality
- topic
- freshness
- early engagement
- text/media quality
- spam risk
Relationship features:
- connection strength
- company affinity
- topic affinity
- past interactions
👉 Interview Answer
Feed ranking usually combines user, item, context, and relationship features. In LinkedIn’s case, relationship strength and professional topic relevance matter more than in a generic entertainment feed.
6️⃣ Quality and Integrity Controls
Controls:
- spam filtering
- low-quality content demotion
- misinformation or policy checks
- duplicate content reduction
- engagement bait detection
- author reputation
7️⃣ Diversity and Freshness
Reranking may avoid:
- too many posts from same author
- too many posts on same topic
- stale content
- low-quality viral content
- over-personalized narrow feed
👉 Interview Answer
The final feed is usually not raw model score order. A reranking layer applies diversity, freshness, policy, and product constraints so the feed feels useful rather than repetitive or low quality.
8️⃣ Feedback Loops
Positive signals:
- dwell time
- reactions
- comments
- shares
- follows
- profile visits
Negative signals:
- hide post
- unfollow
- report
- quick scroll
- low dwell time
9️⃣ Staff-Level Trade-offs
| Decision | Benefit | Cost |
|---|---|---|
| Optimize engagement | Short-term activity | May reduce feed quality |
| Add quality filters | Better trust | Lower raw CTR |
| Use network graph | Professional relevance | Graph fan-out cost |
| Personalization | Higher relevance | Filter bubble risk |
| Freshness boost | Timely feed | Can hurt evergreen quality |
中文部分
中文速记
一句话
LinkedIn Feed 不是纯娱乐推荐，而是多阶段 feed ranking：召回候选内容，再按职业相关性、关系强度、质量和 trust 排序。
背诵要点
- candidate generation 来自 connection、company、creator、group 和 recommendation
- ranking features 包括 user、post、context、relationship
- LinkedIn 更重视 professional relevance 和 trust
- 最终结果不是 raw model score，还要 reranking
- 核心权衡是 engagement vs long-term professional value
中文面试回答
我会把 LinkedIn feed 设计成多阶段排序系统。 第一阶段 candidate generation 从一度联系人、关注公司、关注 creator、group、网络互动内容和推荐内容里召回候选帖子。 第二阶段 feature enrichment 会补充用户行业、职位、技能、关系强度、作者质量、帖子主题、freshness 和历史互动等特征。
排序模型可以预测点击、停留、评论、分享和关注等行为，但最终不能只优化 engagement。 LinkedIn 的目标还包括职业相关性、内容质量、trust and safety、反垃圾和长期用户价值。 所以最终还需要 quality filters、diversity、freshness 和 integrity controls。
Staff 级重点是：专业社区 feed 不能被纯粹的短期点击率主导。 好的设计要在 engagement、trust、内容质量和 professional value 之间做平衡。
✅ Final Interview Answer
A LinkedIn-like feed ranking system uses a multi-stage pipeline. Candidate generation retrieves posts from connections, followed companies, creators, groups, and recommended professional sources. The ranking model then scores candidates using user, post, context, and relationship features such as industry, skills, connection strength, freshness, engagement, and content quality.
The final feed should not be pure engagement ranking. A post-ranking layer applies spam filters, quality controls, diversity, freshness, and professional relevance constraints. At staff level, the key trade-off is balancing short-term engagement with long-term trust and professional value.
Implement