Key Notes
- Billions of videos, Millions of users
- Recommend video for right user / time / context
- Start small, Scale to current needs
- Recommendation based on the entire corpus of content
- Meat in the middle approach
- Pick a Small dataset
- Video, User1, Completion %%
- Video, User2, Completion %%
- Use for a similar taste
- Similarity-based on some model
- The user doesn't like video
- Collaborative filtering techniques
- Activity, Attributes, Demographics
- User - User
- Demographics, Age, locale, Historical data
- Weightage of recent vs previous topics
- Topic Continuity based recommendations
- Reference / Approximation
- Comparing across videos/users
- Auto recommendations / Next list to display
- Map user/item vector and compare with Cartesian steps
- Edge cases - Cold start / new user content
- The initial filter of videos based on preferences then include trending videos in the domain, Add scoring
- Use choices long/short videos
- Ranking step / Filtering Step
My thoughts
- Clustering and they recommending based on cluster they belong
- Custom models for each locale, interests then getting it prioritized
Machine Learning System Design (YouTube Recommendation System)
Key Notes
Multitask ranking system
- Two stage pattern
- Select narrowed down candidates
- Ranking approach to check those candidates
- Funnel Fashion
- Could be SQL query / Watched by users
- Regression to narrow down further
- Finding relevant candidates for users
- Current watched video + context
- 700 users / second / scalable system
- Predict the probability of engagement for a video
- For each value multiple predictions
- Combine all outputs, ensemble them
Keep Thinking!!!
No comments:
Post a Comment