Item Factor
When a recommender decides two albums appeal to the same kind of listener, it does not store that judgement as a sentence. It stores a short list of numbers for each album, and item factors are those numbers.
What the vector contains
An item factor is the row that matrix factorization learns for one catalogue item — commonly 32 to 256 values, none of which is assigned a meaning by hand. Affinity between a user and an item is the dot product of that item's factor vector with the user's, usually plus a user bias and an item bias. Training nudges both sides until those dot products reproduce the interactions that were observed; every unobserved cell then follows from the arithmetic.
Where it comes from and where it goes
Fitting happens offline, by stochastic gradient descent on rating data or by alternating least squares on implicit counts, as the implicit library does on MovieLens-scale data. The resulting item matrix is the artefact that gets served: load it into FAISS and candidate generation becomes a nearest-neighbour lookup against one user vector instead of a scan of the catalogue.
Why it matters
The item factor is where a cold item hurts most. With no interactions there is nothing to fit, so the vector stays near its random initialisation and the item is effectively unrankable until side information supplies a substitute. Two further properties earn the representation its place:
- Item-item similarity becomes a cheap vector comparison, with no co-rating users to intersect.
- The same vector is reusable — it is exactly the embedding a neural recommender starts from.
An item factor is the low-dimensional vector that matrix factorization learns for one catalogue item, typically a few dozen values with no hand-assigned meaning. Scored against a user's factor vector by dot product, it predicts affinity for interactions never observed. The learned item matrix is also the artefact retrieval serves from, so nearest-neighbour search replaces scanning the catalogue. An item with no interaction history has no fitted factor, which is why cold items rank badly.
Think of a record-shop assistant who, after years of watching what sells, could describe any album with a handful of dials: how upbeat it is, how mainstream, how late-night. An item factor is that set of dials — except the software invents the dials itself, purely by watching which people enjoy which things.
Nobody tells it what the dials mean, and nobody needs to. It only needs numbers that let it compare you with an album and guess whether you will like it. Albums nobody has played yet have no dial settings at all, which is why brand-new things are so hard to recommend well.