Item
In ordinary speech an item is just a thing. In recommendation it is one of the two entities the whole problem is built from: the units a system is allowed to recommend, each carrying an identifier that interaction data can be attached to.
What makes something an item
An item is whatever occupies one slot in a recommendation. The choice is a modelling decision rather than a given, and it decides what the system can possibly learn. A music service can treat a track as the item, or an album, or an artist; a marketplace can recommend a product listing or a seller. Choose too fine a unit and every item is starved of data; too coarse and the recommendation stops being actionable — "we suggest this artist" when the reader wanted a song.
How items enter the model
Items form one axis of the user-item matrix, one tower of a two-tower model, and the candidate pool that retrieval narrows. Each acquires an item factor or embedding learned from who interacted with it, plus an item bias absorbing how well liked it is in general. Their frequency distribution is brutally uneven: a few items carry most of the interactions and the long tail carries almost none, which is the fact underneath popularity bias and cold start.
Why the word is deliberately bland
Calling them items rather than films or products is what lets one algorithm serve every domain. The MovieLens datasets, the implicit ALS implementations and every two-tower reference model all speak of users and items, and a method written against that abstraction moves from a video catalogue to a grocery aisle unchanged.
An item is a unit a recommender is able to recommend — a film, a product listing, a track — identified consistently so interactions can be attached to it. Items form one axis of the user-item matrix and the candidate pool that retrieval narrows. The granularity is a modelling decision: too fine and each item has too little data, too coarse and the recommendation is not actionable. Item frequencies are highly skewed, which is what drives popularity bias.
If a shop is going to suggest something to you, it first has to decide what counts as a "something". That unit is what recommendation people call an item.
An item is one thing the system is allowed to put in front of you: a film, a song, a product, a news story. Everything the system learns is tied to these units — who watched this, who bought that. Picking the right size of unit matters more than it sounds. Suggesting a whole album is less useful than naming the one song you would love, but if the units are too small there is rarely enough history about any single one of them to learn anything.