System-Design-Notes
[System Design] Video RCMD System
High-Level Design
Candidate Generation
Scoring
Prioritize accuracy over efficiency
Use conten-based filtering and heavier models relying on video features.
User Features
User-Video Interaction Features
Video Features
Types of RCMD System
Content-based filtering: use video features to recommend videos similar to previous
Handle new videos, slower, need domain knowledge, cannot find new interest
Collaborative filtering (CF): use user-user or video-video similarities
Cannot handle new videos, faster, no domain knowledge, discover new interest
Hybrid filtering: combine CF-based and content-based sequentially or in parallel
Two-Tower Network (candidate generation)
Re-Ranking
- Region-restricted videos
- Videos freshness
- Videos spreading misinformation
- Duplicated
- Fairness and bias
Cold-Start Problem
New users:
- Use user basic features: age, gender…
- After the user interacts with more videos, can do better
New videos:
- Use baisc metadata and content
- Display videos to random users to collect interaction data
Summary
