Item Profile
Before a content-based recommender can decide whether you will like a film, it has to write down what the film actually is — and it has to do that without consulting anybody's viewing history. That written-down form is the item profile.
What goes into it
An item profile is a numeric description built only from an item's own attributes: genre, director and language as one-hot or multi-hot categorical features; a synopsis as TF-IDF weights or a sentence embedding; editorial tags, duration, price, release date. The pieces are concatenated, weighted and normalised into a single content vector that lives in the same space as every other item's profile.
How it is used
The user profile is typically an aggregate of the profiles of items that user liked, so scoring is a cosine similarity between two vectors — profile matching, with the interaction matrix never consulted. That independence is the point: a film uploaded an hour ago has a complete profile from its metadata alone and is immediately rankable, which is the standard structural answer to new-item cold start.
What it cannot do
A profile is only as good as the metadata behind it, and the failure modes follow directly from that:
- Sparse or inconsistent tagging makes unrelated items look alike, and the recommendations inherit the error.
- Attributes capture subject, not quality — two films with identical metadata can be excellent and unwatchable.
- Scores stay inside the attributes a user has already engaged with, which narrows discovery and feeds filter bubbles.
An item profile is a numeric description of a catalogue item built from its own attributes — categorical fields, editorial tags, and text features drawn from a title or synopsis — concatenated into a content vector. Scoring compares it against a user profile aggregated from items that user liked, so no interaction matrix is required. This makes a newly added item rankable the moment its metadata exists, the standard remedy for new-item cold start, and only as reliable as that metadata.
Imagine a librarian filling in an index card for every new book that arrives: subject, author, reading level, a line about what happens inside. She has never watched a single person borrow it, yet she can already tell you who it would suit.
An item profile is that index card, written in numbers so software can compare thousands of cards in an instant. Its great advantage is needing no history. A film added this morning already has a card, so it can be suggested today. The catch is that a card describes what something is about — never whether it is any good.