Slate Recommendation
A homepage row is not one recommendation repeated ten times. It is a single object — a slate — and part of what makes it good is what the items do to each other.
Scoring the list, not the item
Standard top-N recommendation scores every candidate independently and keeps the best k. That framing cannot express substitution (three near-identical thrillers in one row waste the row), complementarity (a camera beside its lens), or position effects (slot one is seen far more than slot six). Slate recommendation makes the whole ordered set the decision variable, so the objective is the expected utility of the list rather than the sum of item scores.
How it is approached
- Greedy re-ranking against a submodular objective — maximal marginal relevance, or a determinantal point process — which keeps per-item scoring and adds an interaction penalty.
- Sequential decoders that emit the slate position by position, conditioning each choice on what is already placed.
- Slate bandits and off-policy estimators that treat the whole slate as the action, which needs a propensity model to learn from logged slates.
Why it matters
The action space is the set of ordered k-subsets of the catalogue, so exact optimisation is hopeless and every practical method approximates over a few hundred retrieved candidates. The framing earns its cost because diversity, freshness and business constraints get enforced coherently instead of being bolted onto a pointwise ranker — and because clicks in a logged slate are confounded by position, which is exactly why item-level offline gains so often fail to move a row's click-through rate.
Slate recommendation treats a whole ordered list — a homepage row, a results page — as a single decision rather than as k independently top-scoring items. The framing captures effects a pointwise ranker cannot express: substitution between near-duplicates, complementarity, and position bias. It is addressed by greedy submodular re-ranking, sequential slate decoders, or slate bandits, and it is where diversity and freshness constraints are enforced coherently instead of patched onto item scores.
A shop window is not ten separate decisions. Put the same jacket in it ten times and you have wasted the window, even if that jacket is your best product.
Slate recommendation is the idea that software should choose a whole row of suggestions as one unit, the way a window dresser arranges a display. What matters is not only how appealing each item is on its own, but whether the items work as a set — whether they cover different moods, whether one makes another more tempting, and which spot each one gets, since the first thing in a row is noticed far more than the last.