🎯Key Takeaways
- 1Collaborative filtering: 'users who bought X also bought Y' — based on user behavior similarity
- 2Content-based filtering: recommend items similar to what the user has liked — based on item features
- 3Two-tower architecture: user tower + item tower → dot product for fast similarity computation
- 4Cold start problem: new users and new items lack data — use content features and popularity as fallback
How Netflix, YouTube, and Amazon Recommend
Recommendation systems drive engagement at every major platform: YouTube's watch time, Amazon's purchases, Spotify's listening, TikTok's scroll time. The core challenge: from millions of items, find the 10-50 most relevant for each user, in real-time.
Modern systems use a multi-stage pipeline: Candidate Generation (narrow millions to hundreds) → Ranking (score and sort) → Re-ranking (diversity, business rules, filters).
Recommendation Approaches
Based on user behavior patterns: users who behaved similarly will like similar things.
User-based: find users similar to you, recommend what they liked.
Item-based: find items similar to what you liked, recommend those.
Limitation: cold start — can't recommend for new users or items with no interaction data.
Advantages
- •Collaborative filtering captures non-obvious user preferences
- •Two-tower architecture scales to billions of items
- •Multi-stage pipeline balances relevance with performance
Disadvantages
- •Cold start requires fallback strategies
- •Feedback loops can create filter bubbles
- •Recommendation quality depends heavily on data quality
🧪 Test Your Understanding
What is the cold start problem in recommendation systems?