Content-Based Recommender

Before a recommender knows anything about who watched what, it already knows what the things are. A content-based recommender runs entirely on that: it proposes items resembling the ones you have already engaged with, judged by the items' own attributes.

How the score is produced
Each item becomes a content vector over its genre, tags, brand or text. The user's engaged items are aggregated into a user profile in that same space, weighted by rating or recency, and each candidate is scored by its similarity to that profile. Notice what is absent: no other user appears anywhere in the computation. The model is per-user and self-contained, which is why it can be as simple as a cosine or as elaborate as a small classifier trained on one person's history.

Why one stays in the stack

New items are scorable immediately, which is the half of cold start collaborative filtering cannot reach.
Nothing has to be popular to be retrievable, so long-tail items stay reachable for users with narrow taste.
Explanations come free: because you watched a Korean thriller is a true statement about the score, not a story told afterwards.
In production it rarely stands alone — it supplies a candidate source and side features to a hybrid or to a model like Wide & Deep.

Its characteristic failure
It cannot surprise anyone. Every suggestion resembles something already chosen, so novelty and serendipity suffer and the recommendations narrow over time. It inherits whatever its metadata lacks — twenty genres cannot separate two thousand dramas. And a brand-new user leaves it with nothing to build a profile from, so user cold start remains untouched.

A content-based recommender scores items by their own attributes against a profile built from the items a user has already engaged with — typically the similarity between a user profile vector and each candidate's content vector. No other user's behaviour enters the computation, so it can rank a brand-new item immediately and explain every choice it makes, at the cost of recommending little beyond what the user has already seen.

Imagine a bookshop assistant who never asks what other customers bought. She just looks at the three books in your hands — same subject, same era, same sort of length — and walks you to the shelf with more like them.

That is a content-based recommender. It describes every item by its own features and suggests things that resemble what you already picked. Its strength is independence: it needs nobody else's opinion, so something that arrived this morning can be recommended this afternoon. Its weakness is predictability. It keeps handing you close neighbours of your own taste, and it will never be the one to show you something you would never have thought to look for.