Personalisation
Two people open the same app on the same morning and see different things. That difference is the entire product of a recommender, and personalisation is the technical name for it: ranking one catalogue differently per user, from what that user has done rather than from what the average user does.
What makes it personalisation and not just relevance
A popularity chart is relevant — it is simply not personalised, because every user receives the same list. Personalisation requires the score to depend on a per-user quantity: a user factor learned from their interactions, an embedding of their recent session, a profile of the attributes they engaged with. Swap the user and the ordering must change, or nothing personal is happening.
Degrees of it
- Segment-level — one list per cohort, such as region or new-versus-returning. Cheap, robust, and the usual fallback.
- Individual — a score per user-item pair, which is what matrix factorisation and two-tower models produce.
- Session-level — the ordering shifts within a single visit as signals arrive, the regime sequential models work in.
Why it is harder than it sounds
Personalisation is bounded by what a user has given you, so a new account has nothing to personalise from — the cold-start problem is the absence of this quantity, not a defect in the model. Pushed hard, it also narrows: a system that only ever extrapolates from past behaviour produces the filter bubble, which is why diversity and novelty are measured alongside accuracy rather than after it.
Personalisation is ranking the same catalogue differently for each user, from a score that depends on a per-user quantity — a learned user factor, a session embedding, or a profile of attributes engaged with. A popularity chart is relevant but not personalised, because every user receives one list. It is the point of a recommender, and it is bounded by how much history a user has: with none, there is nothing to personalise from.
Two people walk into the same bookshop and are handed different books. That is personalisation. The shop's stock has not changed — the choice of what to put in front of you has, based on what you have read before. Streaming services, shops and feeds all work this way: the catalogue is the same for everyone, but the order it is shown in is yours. It only works if the software has something to go on. On your first visit it knows nothing about you, so you get whatever is popular — which is why apps feel generic at first and sharpen as you use them.