Content Vector

Strip away everything a recommender has learned from clicks and the item itself remains — its genre, its tags, its price, its language. A content vector is that description written as a row of numbers, so items can be compared by what they are rather than by who has already engaged with them.

Building the row of numbers
Every item is encoded into the same fixed-length vector: one-hot blocks for single-valued categories, multi-hot or TF-IDF-weighted blocks for tags and text, learned embeddings for high-cardinality fields such as brand, and normalised numeric fields for price or duration. Because all items share that space, the cosine between two vectors is a usable measure of resemblance. A user profile lives in the same space — a recency-weighted average of the vectors of items the user engaged with — which reduces profile matching to a single dot product.

Where it earns its place

• A film uploaded this morning has a complete content vector and zero interactions, so it can be retrieved and scored on day one.
• Concatenated onto a learned item embedding, it gives a factorisation model something to lean on while interaction signal is still thin.
• Indexed in FAISS, the vectors turn a similar-items query into an approximate nearest-neighbour lookup instead of a scan over the catalogue.

What it cannot represent
The vector knows only what was recorded. Thin metadata leaves thousands of items with near-identical vectors, and an unnormalised price field will quietly dominate every similarity it enters. More fundamentally it describes items, not taste, so similarity converges on more of the same — the narrowness content-based filtering is known for.