Machine Learning
Unit 5: Trends and Applications / Prediction & Recommendation
4. Prediction & Recommendation
Recommendation systems predict user preferences to suggest items (movies, products, music) they are likely to engage with. They power platforms like Netflix, Amazon, and Spotify.
Collaborative Filtering
Based on the principle: "Users who agreed in the past will agree in the future." It relies purely on user-item interaction history.
- User-Based: Find users similar to the target user; recommend what they liked.
- Item-Based: Find items similar to what the target user liked.
- Matrix Factorization: Uses SVD/ALS to decompose the rating matrix into latent factors.
Content-Based Filtering
Recommends items similar to ones the user previously liked, based explicitly on the features of the items (genre, tags, actors).
- Builds a specific "user profile" based on item features.
- Uses metrics like Cosine Similarity or TF-IDF.
- Disadvantage: Over-specialization (creates a "filter bubble", no serendipity).
Hybrid Recommendation Systems
Combines Collaborative and Content-Based methods to overcome their individual limitations (e.g., Netflix combines user viewing history with deep tagging of movie genres/actors).
Deep Dive: The Cold Start Problem
The "Cold Start" is the primary weakness of Collaborative Filtering—what happens when there is no data?
| Cold Start Scenario | The Problem | Solutions |
|---|---|---|
| New User | System has zero history on what the user likes. | Ask for preferences at signup (onboarding survey), use broad demographic data. |
| New Item | A newly added movie/product has 0 ratings. | Use Content-Based filtering to push it to users who like its specific features. |
| New System | A brand new app has zero users AND zero ratings. | Web crawling, buying external datasets, or importing social network data. |