00 · IN THREE MINUTES
The answer in three steps
- 1Clicks, viewing time, ratings and skips become imperfect signals of preference.
- 2Models find patterns among users and items, then rank candidates for a specific context.
- 3Recommendations change future behavior, creating a feedback loop between prediction and exposure.
01 · BEHAVIOR BECOMES TRAINING DATA
Behavior becomes training data
A click can mean interest, curiosity or even dislike; absence of a click can mean the item was never shown. Systems combine many signals and usually weight recent, deliberate actions differently from weak exposure.
02 · SIMILARITY CAN BE LEARNED
Similarity can be learned
represents users and items through shared latent factors inferred from interaction patterns. Content features and neural ranking models add information about the item, session and device.
Behavior becomes training data
A click can mean interest, curiosity or even dislike; absence of a click can mean the item was never shown. Systems combine many signals and usually weight recent, deliberate actions differently from weak exposure.
signalsSimilarity can be learned
Collaborative filtering represents users and items through shared latent factors inferred from interaction patterns. Content features and neural ranking models add information about the item, session and device.
candidate setRanking serves an objective
The system may optimize watch time, purchase probability, satisfaction, diversity or a mixture. Changing the objective can change the feed even when the underlying user history is identical.
rankExposure creates the next dataset
Items ranked highly receive more observations, while hidden items remain uncertain. Exploration, randomization and counterfactual evaluation help distinguish genuine preference from the system’s own previous choices.
feedback loop03 · RANKING SERVES AN OBJECTIVE
Ranking serves an objective
The system may optimize watch time, purchase probability, satisfaction, diversity or a mixture. Changing the objective can change the feed even when the underlying user history is identical.
04 · EXPOSURE CREATES THE NEXT DATASET
Exposure creates the next dataset
Items ranked highly receive more observations, while hidden items remain uncertain. Exploration, randomization and counterfactual evaluation help distinguish genuine preference from the system’s own previous choices.
05 · PREDICTION IS NOT MIND READING
Prediction is not mind reading
A ranking score applies to an item in a context and carries uncertainty. It can encode popularity and social bias, and it should not be described as a complete portrait of a person.
06 · SOURCES AND EVIDENCE
Sources and evidence
Claims are linked to foundational papers, standards or the primary study behind the update.
- 01Matrix Factorization Techniques for Recommender SystemsSCHOLARLY REVIEW ↗
Supports a defined mechanism, measurement or evidence boundary in this article.
- 02The Netflix Recommender System: Algorithms, Business Value, and InnovationSCHOLARLY REVIEW ↗
Supports a defined mechanism, measurement or evidence boundary in this article.