wunder beta📱 How Recommendation Algorithms Work
You did not search. A list is already waiting. Traces become a metric, a sparse table, then a retrieve-and-rank funnel — not a mind-reader.
What you’ll learn
- What it actually sawSeparate traces of behavior from the claim that a recommender knows what you like, and see how the training objective shapes the list.Explicit ratings are scarce. Implicit watches and clicks are plentiful and ambiguous. The model learns the metric you train, not an inner taste.
- Filling the blanksRead a sparse user-item table, use neighbors, peel off user and item bias, then see why matrix factorization must not treat missing as zero.The Netflix Prize table was about 99 percent empty. Neighbors lend overlap. Baselines eat a lot of a star. Factorization guesses the rest from observed cells only.
- The list on screenWalk the retrieve-score-re-rank funnel and name cold start and popularity loops as limits of learning from exposure.Nobody scores the whole catalog. Generators retrieve; a heavier model ranks; product rules rewrite. Missing history and extra exposure of hits both distort what you see.
Questions this course answers
What is an implicit feedback signal?
Implicit feedback records behavior — watches, clicks, skips, dwell — and infers possible interest. It is not a direct statement of preference.
Match the event to the kind of signal.
Ratings and thumbs state a preference. Watches and clicks are behavior the system interprets, including clicks on misleading titles.
Why can a click-rate objective produce a worse list than it looks?
Google's scoring lesson and Covington et al. both warn that optimizing clicks promotes deceptive titles. Watch time is a different wish, with its own failure (very long videos).
In the Netflix Prize table Koren describes, about what percent of possible user–movie ratings were missing?
Koren's BellKor write-up: about 99 percent of possible ratings were missing, because a person typically rates only a small portion of 17,770 movies.
Why is treating every empty cell as zero a bad way to factor a recommender matrix?
Google's matrix-factorization page: a full-matrix SVD on a sparse table of mostly blanks is pulled toward zero and generalizes poorly. Fit the observed entries; do not pretend silence is a rating of zero.
In Koren's Titanic example, what is Joe's baseline rating before any taste vector?
3.7 − 0.3 + 0.5 = 3.9. A large part of a raw star is how tough the rater is and how beloved the title is, not the pairing.
Grounded in trusted sources
- [object Object]
- [object Object]
- [object Object]
- [object Object]
- [object Object]
- [object Object]
- [object Object]
- [object Object]
Every Wunder lesson is built from real, reputable sources — never invented.
Related Technology courses
Wunder is a personalized learn-anything platform — tell it any topic and it builds a beautiful, fact-checked course in minutes, with narration, a knowledge check, and a college-style University track.
Browse more Technology courses · All topics · Home
© 2026 Wunder Learning LLC · Terms & Privacy