Data and Label Collection¶
Collaborative Filtering (CF)¶
Relies on user-item interactions to recommend items.
Since users rarely provide explicit ratings, implicit signals are inferred from engagement behaviors.
User Engagement-Based Labels¶
Implicit Label |
Collection Method |
Assumptions & Trade-offs |
|---|---|---|
Clicks |
Count clicks on an item. |
Simple; scalable. Clicking liking (accidental clicks). |
Watch Time / Dwell Time |
Measure time spent on videos/articles. |
Captures engagement depth. Long duration satisfaction (e.g.; passive watching). |
Purchase / Conversion |
Track purchases (e-commerce; rentals; subscriptions). |
Strongest preference signal. Sparse data (only a few items are purchased). |
Add to Cart / Wishlist |
Users mark interest without purchasing. |
Softer preference signal. Users may abandon carts. |
Scrolling & Hovering |
Detect mouse hover time over items. |
Early preference signal. May be unintentional. |
Search Queries & Item Views |
Items viewed after searching for a term. |
Strong relevance signal. Some users browse randomly. |
Negative Feedback & Implicit Dislikes¶
CF Use Case Example: - Spotify uses play count, skip rate, and playlist additions to infer user preferences. - Netflix monitors watch completion rate, rewatches, and early exits for movie recommendations.
Content-Based Filtering (CBF)¶
Session-Based & Short-Term Context Labels¶
Implicit Label |
Collection Method |
Assumptions & Trade-offs |
|---|---|---|
Recent Search Context |
Track evolving search terms. |
Captures short-term needs. Trends change quickly. |
Location-Based Preferences |
Match user location with nearby content. |
Useful for local recommendations. Privacy-sensitive. |
Time of Day / Activity Patterns |
Suggest different items based on morning/evening behavior. |
Improves context relevance. Needs continuous adaptation. |
Self-Supervised Paradigm¶
TODO
Knowledge Graphs for Hybrid Labeling¶
Uses entities and relationships to enhance recommendations.
Social & Community-Based Signals¶
Implicit Label
Collection Method
Assumptions & Trade-offs
Likes / Upvotes
Count “likes” on posts; videos; or comments.
Clear positive feedback. Some users never like items.
Shares / Retweets
Count how often users share content.
Strong endorsement. May share for controversy.
Follows / Subscriptions
Followed creators or product wishlists.
Indicates long-term interest. Users may follow without deep engagement.