What is collaborative filtering?

Collaborative filtering is a personalization technique that predicts what a person will like from the behavior of many other people, rather than from any description of the items themselves. It is the method behind the familiar “people who bought this also bought that,” and it remains one of the most widely used engines for recommendations.

How it works

Collaborative filtering comes in two broad shapes.

The first is item-to-item similarity. The system looks at what tends to be consumed together across the whole population and, for whatever a person is viewing now, surfaces its nearest neighbors. The person barely enters the calculation except as a short list of recent items, which is why much of what feels like “the site knows me” is really “these items go together.”

The second is matrix factorization. Picture a large grid with people down one side and items across the top, and a value in a cell wherever someone rated or acted on something. Factorization decomposes that grid into compact latent vectors for each person and each item, so that combining a person’s vector with an item’s vector estimates how much they would like it. The Netflix Prize (2006 to 2009) made this the standard approach for a time.

Today the factorization is usually done with neural networks. Two-tower models learn one vector tower for users and one for items, so millions of candidates can be retrieved quickly, and sequence models read the order of a person’s recent actions to predict the next one. The math is heavier, but the idea is unchanged: learn from behavior, not from content.

What it is good at

Its great advantage is that it needs no understanding of the items. It does not read the product description, the article text, or the movie’s genre; it works purely from the pattern of who interacted with what. That makes it effective across catalogs too large and too varied to describe by hand, and it is why collaborative filtering scaled commerce and media recommendations in the first place.

Where it struggles

Its weaknesses are built in, and they matter for anyone relying on it.

• Cold start. A new user has no history and a new item has no interactions, so the model has little to go on for either. Fresh catalog items and first-time visitors are where collaborative filtering is weakest, which is one reason it is usually paired with content and context in a hybrid recommender.
• Popularity bias. Because the data is thickest around popular items, the model leans toward them. Left unchecked it keeps surfacing the same hits and starves the long tail, which is part of the homogenization concern covered in personalization and its critics.
• A mostly empty grid. The user-by-item matrix is overwhelmingly blank; even a large platform has observed only a tiny fraction of all possible person-item pairs. This is an understated truth of the field: the model usually knows far less than the marketing implies, and its confidence should be read accordingly.

Where it came from

Collaborative filtering is one of the oldest ideas in personalization. The community-filtering systems of the early 1990s, such as GroupLens, were built on explicit ratings people chose to give, and the technique grew from there through item-to-item methods to the matrix factorization of the Netflix Prize and today’s neural retrieval. That lineage is traced in where personalization comes from.

When to use it

Collaborative filtering is a strong default wherever you have plenty of behavioral data and a catalog too big to curate by hand. It is a poor fit on its own for new users, new items, or cold catalogs, where content-based filtering and contextual methods carry more weight. In practice it is rarely used alone: production recommenders combine it with content and rules, and only measurement can tell you whether it is actually helping a given person.