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.

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.

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.