Recommender systems

A recommender system picks a handful of items from a catalog that may hold millions or billions, under a tight latency budget, for a user who has not typed a query. Netflix, YouTube, Spotify, Amazon, Instagram and news feeds share the same skeleton. They retrieve a shortlist cheaply, score it carefully, then apply business rules before the user sees a few cards.

The hard part is not training a neural net. It is choosing what to optimize, learning from feedback that is mostly silent, keeping popular items from drowning out the rest, and keeping a multi-stage system fresh as the catalog changes. The concepts below build that picture. Ranking covers how the second stage orders the shortlist, and news feed covers the systems side.

The funnel

Candidate generation uses one or more retrievers, such as a two-tower model, co-visitation graphs, trending items or subscriptions. Their outputs are merged and deduplicated. The goal is high recall.
Ranking scores each candidate with many features. The goal is to order the shortlist well.
Re-ranking applies rules: diversity, freshness, creator fairness, age gates, ad insertion and "seen recently" filters.

A homepage usually merges several generators, such as "Continue watching," "Because you watched X," "Trending" and "Friends liked." One giant model that scores the full catalog is clean on paper, but almost never shippable at consumer scale.

Concepts in this section

Read them in order. The first two are the basis for the retrieval step.

Step	Concept	What it answers
1	Implicit feedback and objectives	What do clicks and skips mean, and what should the system optimize?
2	Matrix factorization	How do you turn interactions into vectors and predict affinity?
3	Two-tower retrieval	How do you search millions of items in milliseconds?
4	Cold start and feedback loops	What goes wrong with new items, popular items and the model's own data?
5	YouTube recommendations, 2016	How did one large system put these pieces together?

Common mistakes

Arguing about architecture before the metric. Settle the product goal first. Clicks, watch time and satisfaction lead to different systems.
Reading logs without absences. Impressions, positions and items that were never shown matter as much as clicks.
Assuming retrieval finds everything. Retrieval can miss good items. Invest in recall until online metrics stop moving.

Key takeaways

Recommenders are multi-stage: cheap, high-recall retrieval, then an accurate ranker, then rules and diversity.
Implicit feedback is abundant and noisy. Pick objectives carefully, since watch time is not clicks.
Matrix factorization is the classic baseline. Two-tower models with ANN search are the modern retrieval default.
Cold start, popularity bias and feedback loops often hurt more than the model architecture does.
Large language models mostly improve item and query embeddings and content features inside this funnel. End-to-end generative ranking of the whole catalog is still early.