User-Item Interaction
Every recommender starts from the same raw material: a record that one person did something with one thing. That single event — a play, a purchase, a skip — is a user-item interaction, and it is the atom the whole field is built from.
What counts as an interaction
An interaction is a (user, item) pair carrying whatever the platform observed: a rating, a click, a twelve-second watch, an add-to-cart, a dismissal. Some are explicit — the user deliberately stated a preference. Most are implicit: a by-product of normal use, where engagement stands in for liking and the absence of an event means nothing definite. Interactions carry a timestamp and often a context — device, session, slot position — which is what makes session-based and context-aware models possible.
Why the unit matters
Collaborative filtering has no item attributes to fall back on, so interactions are its only signal; stacked into the user-item matrix they are what factorisation decomposes and what neighbourhood methods compare. The MovieLens datasets ship as little more than interaction tuples, and that is enough to train implicit ALS or BPR.
Where it misleads
An interaction records exposure as much as preference: the user clicked that item because the ranker showed it, so position bias and popularity bias are baked into the data before any model sees it. A brand-new user has no interactions at all — the cold start problem in its simplest form.
A user-item interaction is a single recorded encounter between one user and one item — a rating, click, play, purchase or skip — usually with a timestamp. It is the atomic observation collaborative filtering learns from, and the cell it fills in the user-item matrix. Most interactions are implicit, so they signal engagement rather than stated preference, and they reflect what the system chose to show as much as what the user wanted.
Think of a shop assistant quietly noting every time you pick something off a shelf, put it back, or take it to the till. None of those notes is a conversation about your taste — but together they say a lot.
A user-item interaction is one of those notes: a record that a particular person did something with a particular thing. You watched this video. You skipped that song. You bought that book. Recommendation systems rarely have anyone's taste described to them directly, so they work almost entirely from these small events, piled up by the million, and guess what you would like next.