Latent Factor
Nobody fills in a form saying they prefer slow character studies with unreliable narrators, yet a few hours of viewing history give it away. A latent factor is one coordinate of that unstated taste — a number a model infers from behaviour, never a label anyone supplied.
What the numbers actually are
Matrix factorization gives every user a vector of d latent factors and every item a vector in the same space, then scores a pair by their dot product. The vectors are fitted to the observed cells of the sparse interaction matrix, with d usually between 32 and 256. Nothing forces a factor to mean anything nameable — the space is identifiable only up to rotation — but directions in it do line up with recognisable properties, animation against live action, mainstream against niche.
What the compression buys
- An unobserved pair is scored from the factors alone, so no overlapping raters are needed — the escape from neighbourhood sparsity.
- d numbers stand in for thousands of interactions, so item vectors fit in an index like FAISS and retrieval over millions of items becomes nearest-neighbour search on the dot product.
Where it breaks, and what it became
A brand-new item has no interactions, so its factors are never fitted and it cannot be scored — cold start is structural here, not a tuning problem, which is why attribute-based and hybrid models exist alongside it. The idea itself did not go away: a two-tower retrieval model ends in the same dot product of a user vector and an item vector, with the factors produced by a network over features rather than fitted per id. That is most of what the word embedding means in a modern recommender.
A latent factor is one dimension of a learned, low-dimensional representation of a user or an item, inferred from interaction data rather than taken from any attribute. Matrix factorization places users and items in a shared factor space and predicts preference as the dot product of their vectors, so unobserved pairs are scored from the factors alone. Those vectors are what make retrieval over a huge catalogue tractable, and what neural recommenders generalise into embeddings.
Imagine a shop that works out, without ever asking, that you go for comfortable rather than fashionable. Nobody wrote that on your account. It is a pattern read out of what you bought.
A latent factor is one of those unspoken qualities turned into a number. Software looks at what millions of people chose and invents a small set of hidden dials — maybe ten, maybe a hundred — then gives every person and every item a setting on each dial. Latent simply means hidden: the dials were never labelled, and most of them have no tidy name. To guess whether you will like something, the software checks how closely your settings line up with the item's.