Re-ranking
A ranker scores items one at a time, which means it can hand back ten excellent recommendations that are nearly the same recommendation ten times over. Re-ranking is the stage that looks at the list as a whole and fixes what per-item scoring cannot see.
Why a third stage exists
Retrieval narrows the catalogue to a few hundred candidates; ranking scores each one in isolation; re-ranking then reorders that short list against objectives that only exist at list level. Because it handles tens of items rather than millions, it can afford expensive logic — comparisons between candidates, constraint solving, or a sequential model that conditions each slot on what has already been placed above it.
What gets adjusted
- Diversity: greedy methods such as maximal marginal relevance penalise a candidate for resembling items already chosen.
- Business rules: collapse a series into one entry, cap how many slots a single seller occupies, pin sponsored placements, hold a quota for fresh items.
- Exploration: inject a small number of uncertain items so tomorrow's training data is not purely what the model already believed.
- Calibration: match the genre mix of the list to the mix in the user's own history instead of collapsing onto their single strongest taste.
Why it is load-bearing
Ordering by predicted relevance alone is how popularity bias and filter bubbles reach the product: every item is individually defensible, the list is monotonous, and engagement falls while offline metrics rise. Re-ranking is where a system knowingly trades a little predicted relevance for variety, fairness and freshness — in one place a human can read and change.
Re-ranking is the final stage of a recommender, reordering the short list a ranking model produced to optimise properties of the list rather than of single items — diversity, deduplication, freshness, business rules and injected exploration. Working on tens of candidates makes list-level logic affordable. Without it, per-item relevance scoring yields repetitive, popularity-skewed lists that look strong offline and underperform with real users.
Imagine a playlist built only from your favourite song's closest matches. Every track is a reasonable guess, and yet the whole thing feels like the same song on repeat.
Re-ranking is the final tidy-up that prevents this. Software first picks a shortlist of things you would probably like, then steps back and judges the shortlist as a set: drop the third nearly identical item, leave room for something unfamiliar, avoid five products from one seller. It trades a little of what you are most likely to like for a list that is actually pleasant to look at — the difference between a technically correct selection and a good one.