Recommendation Scenario
The recommender is rarely one thing. A homepage shelf, a strip of related products under an item, and a re-engagement email are three different jobs wearing the same name, and the word for each of them is a recommendation scenario.
What defines one
A scenario is a specific slot in the product together with everything that slot fixes: what triggers it, what context is available, how many items fit, and what counts as success. An item-page strip is seeded by an anchor item and judged on whether a related purchase follows. A homepage rail has no anchor, only the user's history, and is judged on session engagement. A push notification carries room for one item and a real cost for being wrong, because an unwanted interruption can lose the user outright.
Why the distinction is practical
- Candidate sources differ: item-to-item similarity for the anchored slot, personalised retrieval for the open-ended one.
- Objectives differ, so one model ranked by click-through can be right in one slot and wrong in the next.
- Slate size and latency budget differ, and an email assembled overnight permits models no on-page slot could afford.
Where it bites
A model trained on one scenario's logs and deployed in another inherits that scenario's position bias and selection bias, which is why multi-scenario training and per-slot evaluation exist as separate concerns. Offline numbers are comparable only within a scenario; a system-wide average hides the slot that is quietly failing.
A recommendation scenario is a specific placement where recommendations are served — a homepage rail, a related-items strip on an item page, a notification — together with its trigger, available context, slate size and success metric. The distinction is operational: each scenario needs its own candidate sources, objective and latency budget, and logs from one carry biases that make a model trained on them unreliable in another.
The suggestions a service shows you are not all produced by the same machinery. The row on the home screen, the strip of related products beneath an item, and the email that arrives on a Friday are separate jobs.
Each of those slots is a recommendation scenario. What matters changes with the slot: beneath a product, suggestions should relate to the thing you are looking at; on the home screen, they should reflect your taste in general; in an email, a bad guess is far more annoying, so the bar is higher. Naming each slot separately keeps a system honest about which one is working and which one is not.