User Profile (Content)

A content-based recommender never looks at what anybody else did. It knows what items are made of, and it knows what you have engaged with — so it describes you in the very vocabulary it uses to describe the catalogue.

How the profile is built
Each item you consumed is already a content vector: TF-IDF weights over its text, one-hot genres, tags, or a learned embedding. The profile is an aggregate of those vectors — commonly their centroid, weighted by rating or recency, with disliked items subtracted — or, equivalently, the weight vector of a small classifier trained on your own positives and negatives. Scoring a candidate is then a cosine similarity between the profile and that item's vector, with no other user involved at any point.

What this buys

• A brand-new item is recommendable the moment its attributes exist — the new-item cold start case that collaborative filtering has no answer to.
• Only one user's own history is needed, so someone with unusual taste is served as well as a mainstream user.
• The profile is readable — mostly long-form documentary and political non-fiction — so an explanation can be generated straight from it.

Where it narrows
A profile can only point toward features it already contains, so the recommendations converge on more of the same: no serendipity, and nothing the feature vocabulary cannot express — humour, craft, why a book is worth finishing — ever reaches the score. Drifting taste needs an explicit recency decay or a sliding window. That narrowing is the standard argument for a hybrid, keeping content profiles for cold items and collaborative signal for everything else.