Content-Based Filtering

Content-based filtering is a recommender-system approach that suggests to each user items similar to those they already liked, by analyzing the items' own attributes. Its premise is straightforward: if you liked something, you will probably enjoy things with similar characteristics. Unlike collaborative filtering, it does not use information about other users: each recommendation rests only on your history and on the properties of the content.

How it works

The method has three pieces. First, it represents each item by its features—genre, keywords, author…—; in the case of text, it usually turns them into vectors with techniques such as TF-IDF or with embeddings. Second, it builds a user profile by aggregating the features of the items the user liked. And third, it recommends by similarity: it measures how much each unseen item resembles the user's profile, usually with cosine similarity, and offers the most akin ones.

Versus collaborative filtering

It is worth contrasting it with collaborative filtering, which recommends from the behavior of many users (“people similar to you also liked this”). The content-based one looks inward, at the item's attributes; the collaborative one, outward, at the community. In practice, real systems usually combine both.

Its advantages

It has clear strengths. It is independent of other users, which makes it easy to scale; it can recommend new items from the very first moment, as long as their attributes are known, without suffering the cold start that afflicts the collaborative approach; it captures niche tastes; and it offers transparent recommendations, easy to justify (“because you liked works of genre X”).

Its limits

It also has weak points. It tends toward over-specialization: since it only recommends what is similar to the already known, it encloses the user in a “bubble” and offers little surprise or novelty. It suffers the cold start of the new user, of whom there is no history yet. And it depends entirely on the existence of good features for the items, a task that often has to be done by hand.