User (RecSys)

In a recommender, the word user stops meaning a person and starts meaning a row. Whatever the system knows about you, it knows as one line of the interaction matrix: the items that line touches, and how strongly it touches each one.

What the row actually is
A user is an identity the system can attach history to — an account, a device, a cookie, sometimes just the current session — and it is the unit everything is personalised to. The representation changes with the model family: a raw interaction vector in neighbourhood methods, a learned latent factor vector in matrix factorisation, the output of a user tower in two-tower retrieval. Side information such as country or signup date hangs off the same identity.

Where the abstraction leaks

• One household account is several people in a single row, which is why a family profile produces a confused mix of recommendations.
• A logged-out person on three devices becomes three users, each with a thin history.
• A freshly created identity has an empty row — the new-user cold start case, where the system has nothing personal to work from and must fall back on popularity or stated preferences.

Why the definition matters
The user is the aggregation unit for evaluation as well as for training: hit rate and NDCG are averaged per user, leave-one-out and temporal splits are taken per user, and the sparse half of the user-item matrix is counted per user. Pick the wrong identity key and both the learning signal and every number used to judge it are quietly wrong.